diff --git a/shrink/llava-v1.6-mistral-7b-hf-mic.html b/shrink/llava-v1.6-mistral-7b-hf-mic.html index 6cd323498ab901a1566bf6227b04956c0776a39b..1db4aeab016c77bf3ae1eb136f58d4c30571c10f 100644 --- a/shrink/llava-v1.6-mistral-7b-hf-mic.html +++ b/shrink/llava-v1.6-mistral-7b-hf-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

llava-hf/llava-v1.6-mistral-7b-hf

huggingface-mic/llava-v1.6-mistral-7b-hf-mic@2575b1c20c76 ← llava-hf/llava-v1.6-mistral-7b-hf@2424fdd47412 · LlavaNextForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original7.6Bparameters
Fixture4.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.vision_tower.encoder.layers

Original · 24
Fixture · 2
KindOriginalFixture
CLIPEncoderLayer242✓ kept

model.language_model.layers

Original · 32
Fixture · 2
KindOriginalFixture
MistralDecoderLayer322✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.language_model.layers7.0B92.2%74.0K1.7%
model.vision_tower.encoder.layers302.3M4.0%66.9K1.5%
model.language_model131.3M1.7%2.1M47.4%
lm_head131.3M1.7%2.1M47.4%
model.multi_modal_projector21.0M0.3%8.3K0.2%
model.vision_tower1.2M0.0%74.9K1.7%
model4.1K0.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typellava_nextllava_next✓ kept
modelvocab_size3206432064✓ kept
modeltie_word_embeddingsFalseFalse✓ kept
vision_configmodel_typeclip_vision_modelclip_vision_model✓ kept
vision_configvocab_size3200032000✓ kept
vision_confighidden_actquick_geluquick_gelu✓ kept
vision_configimage_size336336✓ kept
vision_configpatch_size1414✓ kept
vision_confignum_channels33✓ kept
text_configattention styleGQA (32 heads, 8 KV)GQA (4 heads, 2 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typemistralmistral✓ kept
text_configvocab_size3206432064✓ kept
text_configtie_word_embeddingsFalseFalse✓ kept
text_confighidden_actsilusilu✓ kept
text_configmax_position_embeddings3276832768✓ kept

Module classes

Instances of each class in the built model (prefix LlavaNext omitted).

ClassOriginalFixture
Linear37129
MistralRMSNorm655
LayerNorm506
MistralAttention322
MistralDecoderLayer322
MistralMLP322
SiLUActivation322
CLIPAttention242
CLIPEncoderLayer242
CLIPMLP242
QuickGELUActivation242
Embedding22
ModuleList22
CLIPEncoder11
CLIPVisionEmbeddings11
CLIPVisionModel11
Conv2d11
GELUActivation11
ForConditionalGeneration11
Model11
MultiModalProjector11
MistralModel11
MistralRotaryEmbedding11

Config changes

11 fields changed, 60 unchanged.

FieldOriginalFixture
text_config.head_dim12816
text_config.hidden_size409664
text_config.intermediate_size14336128
text_config.num_attention_heads324
text_config.num_hidden_layers322
text_config.num_key_value_heads82
vision_config.hidden_size102464
vision_config.intermediate_size4096128
vision_config.num_attention_heads164
vision_config.num_hidden_layers242
vision_config.projection_dim76864
\ No newline at end of file diff --git a/shrink/long-t5-tglobal-base-mic.html b/shrink/long-t5-tglobal-base-mic.html index ddf97b8ed191ad92edc2c4eb1c9f19befaf24c5e..a5b321b1541a05a2b5a406a26add38bffa7416f3 100644 --- a/shrink/long-t5-tglobal-base-mic.html +++ b/shrink/long-t5-tglobal-base-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/long-t5-tglobal-base

huggingface-mic/long-t5-tglobal-base-mic@04d69ed32e31 ← google/long-t5-tglobal-base@aecb1376e5bd · LongT5ForConditionalGeneration · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original222.9Mparameters
Fixture2.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

encoder.block

Original · 12
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding111✓ kept

decoder.block

Original · 12
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding111✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
decoder.block113.3M50.8%115.2K5.1%
encoder.block85.0M38.1%82.6K3.7%
shared24.7M11.1%2.1M91.2%
decoder.final_layer_norm7680.0%640.0%
encoder.final_layer_norm7680.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typelongt5longt5✓ kept
modelvocab_size3212832128✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix LongT5 omitted).

ClassOriginalFixture
Linear21737
Dropout8616
LayerNorm7414
ModuleList266
Attention244
Block244
DenseGatedActDense244
LayerFF244
NewGELUActivation244
LayerCrossAttention122
LayerSelfAttention122
LayerTransientGlobalSelfAttention122
TransientGlobalAttention122
Embedding66
Stack22
ForConditionalGeneration11

Config changes

6 fields changed, 31 unchanged.

FieldOriginalFixture
d_ff2048128
d_kv6416
d_model76864
num_decoder_layers122
num_heads124
num_layers122
\ No newline at end of file diff --git a/shrink/longformer-base-4096-mic.html b/shrink/longformer-base-4096-mic.html index 035aa05d5b555ee106b24375ca6cb4d7cb872cfb..b6ed2f1c8417b84e27ab4b97d0be0818efd3f3ab 100644 --- a/shrink/longformer-base-4096-mic.html +++ b/shrink/longformer-base-4096-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

allenai/longformer-base-4096

huggingface-mic/longformer-base-4096-mic@5e8ae916a995 ← allenai/longformer-base-4096@301e6a42cb0d · LongformerForMaskedLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original148.7Mparameters
Fixture3.6Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

longformer.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
longformer.encoder.layer106.3M71.5%91.9K2.5%
longformer.embeddings41.8M28.1%3.5M96.0%
lm_head.dense590.6K0.4%4.2K0.1%
lm_head50.3K0.0%50.3K1.4%
lm_head.layer_norm1.5K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typelongformerlongformer✓ kept
modelvocab_size5026550265✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings40984098✓ kept

Module classes

Instances of each class in the built model (prefix Longformer omitted).

ClassOriginalFixture
Linear11020
LayerNorm266
Dropout255
GELUActivation122
Attention122
Intermediate122
Layer122
Output122
SelfAttention122
SelfOutput122
Embedding33
Embeddings11
Encoder11
ForMaskedLM11
LMHead11
Model11
ModuleList11

Config changes

5 fields changed, 26 unchanged.

FieldOriginalFixture
attention_window12 entries: 12× 512[512, 512]
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/luke-base-mic.html b/shrink/luke-base-mic.html index a7c2113ff9770b610bea6f81aca31cd2d15e8683..68d2c1c88f45ed8df9c18ca4cd2243204e598bec 100644 --- a/shrink/luke-base-mic.html +++ b/shrink/luke-base-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

studio-ousia/luke-base

huggingface-mic/luke-base-mic@39d0a41c62c2 ← studio-ousia/luke-base@7438924defd9 · LukeForMaskedLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original314.4Mparameters
Fixture135.2Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

luke.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
luke.entity_embeddings128.6M40.9%128.0M94.7%
luke.encoder.layer106.3M33.8%91.9K0.1%
luke.embeddings39.0M12.4%3.3M2.4%
lm_head.decoder38.6M12.3%3.2M2.4%
lm_head.dense590.6K0.2%4.2K0.0%
luke.pooler590.6K0.2%4.2K0.0%
entity_predictions500.0K0.2%500.0K0.4%
3 others249.2K0.1%67.5K0.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typelukeluke✓ kept
modelvocab_size5026750267✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings514514✓ kept

Module classes

Instances of each class in the built model (prefix Luke omitted).

ClassOriginalFixture
Linear11424
Dropout388
LayerNorm288
GELUActivation133
Attention122
Intermediate122
Layer122
Output122
SelfAttention122
SelfOutput122
Embedding66
CrossEntropyLoss11
EntityPredictionHead11
EntityPredictionHeadTransform11
Embeddings11
Encoder11
EntityEmbeddings11
ForMaskedLM11
LMHead11
Model11
Pooler11
ModuleList11
Tanh11

Config changes

4 fields changed, 29 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7-mic.html b/shrink/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7-mic.html index ec2b7e7b0227088f4fc95da049f21d1818970480..bf1651c7c0199052d494ec4d0e20aa04ac03aa6a 100644 --- a/shrink/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7-mic.html +++ b/shrink/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7

huggingface-mic/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7-mic@b088660049cb ← MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7@b5113eb38ab6 · DebertaV2ForSequenceClassification · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original278.8Mparameters
Fixture16.2Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

deberta.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
deberta.embeddings192.8M69.1%16.1M99.4%
deberta.encoder.layer85.1M30.5%66.9K0.4%
pooler.dense590.6K0.2%4.2K0.0%
deberta.encoder394.8K0.1%32.9K0.2%
classifier2.3K0.0%1950.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typedeberta-v2deberta-v2✓ kept
modelvocab_size251000251000✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings512512✓ kept

Module classes

Instances of each class in the built model (prefix DebertaV2 omitted).

ClassOriginalFixture
Linear7414
Dropout5111
LayerNorm266
Attention122
Intermediate122
Layer122
Output122
SelfOutput122
DisentangledSelfAttention122
GELUActivation122
Embedding22
ContextPooler11
Embeddings11
Encoder11
ForSequenceClassification11
Model11
ModuleList11

Config changes

5 fields changed, 31 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
pooler_hidden_size76864
\ No newline at end of file diff --git a/shrink/mamba-130m-hf-mic.html b/shrink/mamba-130m-hf-mic.html index c17b34629362584c80bb8d1d603af0319335d6cc..ea7c00b3b3932bdf38807f99e06208ded08b81d7 100644 --- a/shrink/mamba-130m-hf-mic.html +++ b/shrink/mamba-130m-hf-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

state-spaces/mamba-130m-hf

huggingface-mic/mamba-130m-hf-mic@049c3bff1eb7 ← state-spaces/mamba-130m-hf@1e76775f628f · MambaForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original129.1Mparameters
Fixture3.3Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

backbone.layers

Original · 24
Fixture · 2
KindOriginalFixture
Block242✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
backbone.layers90.5M70.1%87.9K2.7%
backbone.embeddings38.6M29.9%3.2M97.3%
backbone.norm_f7680.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemambamamba✓ kept
modelvocab_size5028050280✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actsilusilu✓ kept

Module classes

Instances of each class in the built model (prefix Mamba omitted).

ClassOriginalFixture
Linear979
RMSNorm253
Conv1d242
Block242
Mixer242
SiLUActivation242
Embedding11
ForCausalLM11
Model11
ModuleList11

Config changes

5 fields changed, 38 unchanged.

FieldOriginalFixture
d_model76864
hidden_size76864
intermediate_size1536128
n_layer242
num_hidden_layers242
\ No newline at end of file diff --git a/shrink/mask2former-swin-tiny-coco-instance-mic.html b/shrink/mask2former-swin-tiny-coco-instance-mic.html index 11167c59e4a664ed766099af7a9b7e042003cd31..fa56e9b3f5cd6a6a118464b176dc8fa380c35780 100644 --- a/shrink/mask2former-swin-tiny-coco-instance-mic.html +++ b/shrink/mask2former-swin-tiny-coco-instance-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/mask2former-swin-tiny-coco-instance

huggingface-mic/mask2former-swin-tiny-coco-instance-mic@851c4b7fb35c ← facebook/mask2former-swin-tiny-coco-instance@22c4a2f15dc8 · Mask2FormerForUniversalSegmentation · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original47.4Mparameters
Fixture8.0Mparameters
Layer stacks4compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.pixel_level_module.encoder.swin.encoder.layers

Original · 4
Fixture · 4
KindOriginalFixture
SwinPatchMerging33✓ kept
no SwinPatchMerging11✓ kept

model.pixel_level_module.decoder.input_projections

Original · 3
Fixture · 3
KindOriginalFixture
Conv2d33✓ kept

model.pixel_level_module.decoder.encoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
FormerPixelDecoderEncoderLayer62✓ kept

model.transformer_module.decoder.layers

Original · 9
Fixture · 1
KindOriginalFixture
FormerMaskedAttentionDecoderLayer91✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.pixel_level_module.encoder.swin.encoder.layers27.5M58.0%4.9M60.7%
model.transformer_module.decoder.layers14.2M30.0%659.3K8.2%
model.pixel_level_module.decoder.encoder.layers4.4M9.3%1.4M17.3%
model.pixel_level_module691.3K1.5%680.0K8.4%
model.pixel_level_module.decoder.input_projections346.4K0.7%231.7K2.9%
model.transformer_module249.9K0.5%191.0K2.4%
class_predictor20.8K0.0%10.4K0.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemask2formermask2former✓ kept
backbone_configmodel_typeswinswin✓ kept
backbone_configtie_word_embeddingsTrueTrue✓ kept
backbone_confighidden_actgelugelu✓ kept
backbone_configimage_size224224✓ kept
backbone_configpatch_size44✓ kept
backbone_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Mask2 omitted).

ClassOriginalFixture
Linear16949
LayerNorm7325
Dropout135
GELUActivation124
ReLU124
SwinAttention124
SwinLayer124
SwinMLP124
SwinRelativePositionBias124
SwinDropPath113
FormerAttention91
FormerMaskedAttentionDecoderLayer91
ModuleList99
MultiheadAttention91
NonDynamicallyQuantizableLinear91
Sequential85
Conv2d710
FormerPixelDecoderEncoderLayer62
FormerPixelDecoderEncoderMultiscaleDeformableAttention62
GroupNorm55
SwinStage44
Embedding33
FormerPredictionBlock33
SwinPatchMerging33
Identity22
FormerSinePositionEmbedding22
FormerForUniversalSegmentation11
FormerHungarianMatcher11
FormerLoss11
FormerMLPPredictionHead11
FormerMaskPredictor11
FormerMaskedAttentionDecoder11
FormerModel11
FormerPixelDecoder11
FormerPixelDecoderEncoderOnly11
FormerPixelLevelModule11
FormerTransformerModule11
ModuleDict11
SwinBackbone11
SwinEmbeddings11
SwinEncoder11
SwinModel11
SwinPatchEmbeddings11

Config changes

9 fields changed, 78 unchanged.

FieldOriginalFixture
backbone_config.depths[2, 2, 6, 2][1, 1, 1, 1]
backbone_config.embed_dim9664
backbone_config.hidden_size76864
backbone_config.num_heads[3, 6, 12, 24][2, 4, 8, 16]
backbone_config.num_layers42
decoder_layers102
encoder_layers62
hidden_dim256128
num_attention_heads84
\ No newline at end of file diff --git a/shrink/mbart-large-50-mic.html b/shrink/mbart-large-50-mic.html index 1a25c0a57b7ce1b58dd2950f822f3bc31cb7f1c7..3604ed0ee3d457e5ff0a5e57107ee27ea194fd91 100644 --- a/shrink/mbart-large-50-mic.html +++ b/shrink/mbart-large-50-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/mbart-large-50

huggingface-mic/mbart-large-50-mic@ef21bcc00bb0 ← facebook/mbart-large-50@4ef55a20b36c · MBartForConditionalGeneration · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original610.9Mparameters
Fixture16.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

model.decoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
DecoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.shared256.1M41.9%16.0M98.2%
model.decoder.layers201.6M33.0%100.5K0.6%
model.encoder.layers151.2M24.7%66.9K0.4%
model.encoder1.1M0.2%65.9K0.4%
model.decoder1.1M0.2%65.9K0.4%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typembartmbart✓ kept
modelvocab_size250054250054✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings10241024✓ kept

Module classes

Instances of each class in the built model (prefix MBart omitted).

ClassOriginalFixture
Linear19333
LayerNorm6414
Attention366
GELUActivation244
DecoderLayer122
EncoderLayer122
ScaledWordEmbedding33
LearnedPositionalEmbedding22
ModuleList22
Decoder11
Encoder11
ForConditionalGeneration11
Model11

Config changes

7 fields changed, 38 unchanged.

FieldOriginalFixture
d_model102464
decoder_attention_heads164
decoder_ffn_dim4096128
decoder_layers122
encoder_attention_heads164
encoder_ffn_dim4096128
encoder_layers122
\ No newline at end of file diff --git a/shrink/mistral3-mic.html b/shrink/mistral3-mic.html index bd26f00b6226745ac099f04496801db17e7c7ed2..ff559eebd0a8201f59ead8838c6ecb4a59984a20 100644 --- a/shrink/mistral3-mic.html +++ b/shrink/mistral3-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

mistralai/Mistral-Small-3.1-24B-Instruct-2503

huggingface-mic/mistral3-mic@ad30ba063d8b ← mistralai/Mistral-Small-3.1-24B-Instruct-2503@68faf511d618 · Mistral3ForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 24 failed checks in 30 days · MIC failures dashboard

Original23.3Bparameters
Fixture8.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.vision_tower.transformer.layers

Original · 24
Fixture · 2
KindOriginalFixture
PixtralAttentionLayer242✓ kept

model.language_model.layers

Original · 40
Fixture · 2
KindOriginalFixture
MistralDecoderLayer402✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.language_model.layers22.2B95.2%74.0K0.9%
model.language_model671.1M2.9%8.4M97.5%
model.vision_tower.transformer.layers402.7M1.7%82.2K1.0%
model.multi_modal_projector35.7M0.2%24.6K0.3%
model.vision_tower603.1K0.0%37.7K0.4%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemistral3mistral3✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
vision_configRoPE typeaxialaxial✓ kept
vision_configmodel_typepixtralpixtral✓ kept
vision_confighidden_actsilusilu✓ kept
vision_configimage_size15401540✓ kept
vision_configpatch_size1414✓ kept
vision_confignum_channels33✓ kept
text_configattention styleGQA (32 heads, 8 KV)GQA (4 heads, 2 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typemistralmistral✓ kept
text_configvocab_size131072131072✓ kept
text_configtie_word_embeddingsFalseFalse✓ kept
text_confighidden_actsilusilu✓ kept
text_configmax_position_embeddings131072131072✓ kept

Module classes

Instances of each class in the built model (prefix Mistral3 omitted).

ClassOriginalFixture
Linear45232
MistralRMSNorm815
SiLUActivation644
PixtralRMSNorm495
MistralAttention402
MistralDecoderLayer402
MistralMLP402
PixtralAttention242
PixtralAttentionLayer242
PixtralMLP242
ModuleList22
Conv2d11
Embedding11
GELUActivation11
ForConditionalGeneration11
Model11
MultiModalProjector11
PatchMerger11
RMSNorm11
MistralModel11
MistralRotaryEmbedding11
PixtralTransformer11
PixtralVisionModel11
PixtralVisionRotaryEmbedding11

Config changes

11 fields changed, 53 unchanged.

FieldOriginalFixture
text_config.head_dim12816
text_config.hidden_size512064
text_config.intermediate_size32768128
text_config.num_attention_heads324
text_config.num_hidden_layers402
text_config.num_key_value_heads82
vision_config.head_dim6416
vision_config.hidden_size102464
vision_config.intermediate_size4096128
vision_config.num_attention_heads164
vision_config.num_hidden_layers242
\ No newline at end of file diff --git a/shrink/mm_grounding_dino_tiny_o365v1_goldg_v3det-mic.html b/shrink/mm_grounding_dino_tiny_o365v1_goldg_v3det-mic.html index f329a7661a54a85d90b921dd4c0cf7a51e75a8e4..39bd520ad1545d2b1ab711657080c002d1ceb4b0 100644 --- a/shrink/mm_grounding_dino_tiny_o365v1_goldg_v3det-mic.html +++ b/shrink/mm_grounding_dino_tiny_o365v1_goldg_v3det-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

openmmlab-community/mm_grounding_dino_tiny_o365v1_goldg_v3det

huggingface-mic/mm_grounding_dino_tiny_o365v1_goldg_v3det-mic@3b6c70187a27 ← openmmlab-community/mm_grounding_dino_tiny_o365v1_goldg_v3det@ef2053aa1e0c · MMGroundingDinoForObjectDetection · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original172.9Mparameters
Fixture7.7Mparameters
Layer stacks7compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.backbone.conv_encoder.model.swin.encoder.layers

Original · 4
Fixture · 4
KindOriginalFixture
SwinPatchMerging33✓ kept
no SwinPatchMerging11✓ kept

model.input_proj_vision

Original · 4
Fixture · 4
KindOriginalFixture
Conv2d44✓ kept

model.text_backbone.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
BertLayer122✓ kept

model.encoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
FusionLayer62✓ kept

model.decoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
DecoderLayer62✓ kept

model.decoder.bbox_embed

Original · 6
Fixture · 2
KindOriginalFixture
Linear62✓ kept

bbox_embed

Original · 6
Fixture · 2
KindOriginalFixture
Linear62✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.text_backbone.encoder.layer85.1M49.2%66.9K0.9%
model.backbone.conv_encoder.model.swin.encoder.layers27.5M15.9%4.9M63.3%
model.text_backbone23.8M13.8%2.0M25.7%
model.encoder.layers21.9M12.7%176.4K2.3%
model.decoder.layers10.9M6.3%142.3K1.8%
model.input_proj_vision2.1M1.2%353.0K4.6%
model.decoder.bbox_embed795.7K0.5%17.2K0.2%
9 others834.0K0.5%93.5K1.2%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemm-grounding-dinomm-grounding-dino✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
backbone_configmodel_typeswinswin✓ kept
backbone_confighidden_actgelugelu✓ kept
backbone_configimage_size224224✓ kept
backbone_configpatch_size44✓ kept
backbone_confignum_channels33✓ kept
text_configmodel_typebertbert✓ kept
text_configvocab_size3052230522✓ kept
text_configtie_word_embeddingsTrueTrue✓ kept
text_confighidden_actgelugelu✓ kept
text_configmax_position_embeddings512512✓ kept

Module classes

Instances of each class in the built model (prefix MMGroundingDino omitted).

ClassOriginalFixture
Linear382122
LayerNorm11943
Dropout6818
ModuleList2719
GELUActivation246
MultiheadAttention186
ReLU186
MLPPredictionHead146
ContrastiveEmbedding135
BertAttention122
BertIntermediate122
BertLayer122
BertOutput122
BertSelfAttention122
BertSelfOutput122
MultiscaleDeformableAttention124
MultiScaleDeformableAttention124
SwinAttention124
SwinLayer124
SwinMLP124
SwinRelativePositionBias124
SwinDropPath113
BiMultiHeadAttention62
DecoderLayer62
DeformableLayer62
DropPath62
EncoderLayer62
FusionLayer62
TextEnhancerLayer62
Conv2d55
Embedding44
GroupNorm44
Sequential44
SwinStage44
SwinPatchMerging33
BertEmbeddings11
BertEncoder11
BertModel11
Identity11
ConvEncoder11
ConvModel11
Decoder11
Encoder11
ForObjectDetection11
Model11
SinePositionEmbedding11
ModuleDict11
SwinBackbone11
SwinEmbeddings11
SwinEncoder11
SwinModel11
SwinPatchEmbeddings11

Config changes

16 fields changed, 93 unchanged.

FieldOriginalFixture
backbone_config.depths[2, 2, 6, 2][1, 1, 1, 1]
backbone_config.embed_dim9664
backbone_config.hidden_size76864
backbone_config.num_heads[3, 6, 12, 24][2, 4, 8, 16]
backbone_config.num_layers42
d_model25664
decoder_attention_heads84
decoder_ffn_dim2048128
decoder_layers62
encoder_attention_heads84
encoder_ffn_dim2048128
encoder_layers62
text_config.hidden_size76864
text_config.intermediate_size3072128
text_config.num_attention_heads124
text_config.num_hidden_layers122
\ No newline at end of file diff --git a/shrink/mms-tts-eng-mic.html b/shrink/mms-tts-eng-mic.html index 9a3f9fc843552b343d8b297aee979f870e9cc220..870fa1548b35c176e7f9d2f59b658e70e8569985 100644 --- a/shrink/mms-tts-eng-mic.html +++ b/shrink/mms-tts-eng-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/mms-tts-eng

huggingface-mic/mms-tts-eng-mic@9594dd28deb2 ← facebook/mms-tts-eng@c71de0fe7204 · VitsModel · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original36.3Mparameters
Fixture16.3Mparameters
Layer stacks5compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

text_encoder.encoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
EncoderLayer62✓ kept

flow.flows

Original · 4
Fixture · 4
KindOriginalFixture
ResidualCouplingLayer44✓ kept

decoder.resblocks

Original · 12
Fixture · 12
KindOriginalFixture
HifiGanResidualBlock1212✓ kept

posterior_encoder.wavenet.in_layers

Original · 16
Fixture · 16
KindOriginalFixture
ModuleDict1616✓ kept

posterior_encoder.wavenet.res_skip_layers

Original · 16
Fixture · 16
KindOriginalFixture
ModuleDict1616✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
decoder.resblocks11.0M30.2%11.0M67.2%
flow.flows7.1M19.6%827.5K5.1%
text_encoder.encoder.layers6.2M17.1%133.1K0.8%
posterior_encoder.wavenet.in_layers5.9M16.3%659.5K4.0%
decoder.upsampler2.7M7.3%2.7M16.3%
posterior_encoder.wavenet.res_skip_layers1.2M3.2%130.9K0.8%
decoder.conv_pre688.6K1.9%688.6K4.2%
13 others1.6M4.3%254.4K1.6%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typevitsvits✓ kept
modelvocab_size3838✓ kept
modelhidden_actrelurelu✓ kept
modelsampling_rate1600016000✓ kept

Module classes

Instances of each class in the built model (prefix Vits omitted).

ClassOriginalFixture
Conv1d177169
ModuleList8080
LayerNorm7264
ModuleDict6464
ParametrizationList6464
ParametrizedConv1d6464
_WeightNorm6464
Dropout2719
Linear248
HifiGanResidualBlock1212
DilatedDepthSeparableConv1010
ConvFlow88
ReLU62
Attention62
EncoderLayer62
FeedForward62
WaveNet55
ConvTranspose1d44
ResidualCouplingLayer44
ElementwiseAffine22
Embedding11
Encoder11
HifiGan11
Model11
PosteriorEncoder11
ResidualCouplingBlock11
StochasticDurationPredictor11
TextEncoder11

Config changes

4 fields changed, 50 unchanged.

FieldOriginalFixture
ffn_dim768128
hidden_size19264
num_attention_heads24
num_hidden_layers62
\ No newline at end of file diff --git a/shrink/mobilenet_v2_1.0_224-mic.html b/shrink/mobilenet_v2_1.0_224-mic.html index 237a1600e30e4f00b96a0fa14237aa0ed475dd58..4dba911e98301907d7670e0e06c75309dfbf7b9d 100644 --- a/shrink/mobilenet_v2_1.0_224-mic.html +++ b/shrink/mobilenet_v2_1.0_224-mic.html @@ -31,6 +31,7 @@ tr.bad td{color:var(--warn)} .c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}

← MIC coverage of Transformers model types

-

google/mobilenet_v2_1.0_224

huggingface-mic/mobilenet_v2_1.0_224-mic@91fe07b79191 ← google/mobilenet_v2_1.0_224@75e607b00aea · MobileNetV2ForImageClassification · overfit weights

-
Original3.5Mparameters
Fixture3.5Mparameters
Layer stacks1compared
Findings1to review

⚠ To review

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

mobilenet_v2.layer

Original · 16
Fixture · 16
KindOriginalFixture
ConvLayer1616✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
mobilenet_v2.layer1.8M51.6%1.8M51.6%
classifier1.3M36.6%1.3M36.6%
mobilenet_v2.conv_1x1412.2K11.8%412.2K11.8%
mobilenet_v2.conv_stem1.8K0.1%1.8K0.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemobilenet_v2mobilenet_v2✓ kept
modelhidden_actrelu6relu6✓ kept
modelimage_size224224✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix MobileNetV2 omitted).

ClassOriginalFixture
BatchNorm2d5252
Conv2d5252
ConvLayer5252
ReLU63535
InvertedResidual1616
AdaptiveAvgPool2d11
Dropout11
Linear11
ForImageClassification11
Model11
Stem11
ModuleList11

Config changes

1 fields changed, 23 unchanged.

FieldOriginalFixture
problem_typenullsingle_label_classification
+

google/mobilenet_v2_1.0_224

huggingface-mic/mobilenet_v2_1.0_224-mic@6938bb7a591f ← google/mobilenet_v2_1.0_224@75e607b00aea · MobileNetV2ForImageClassification · real weights

+

CI: no failure on main in 30 days · MIC failures dashboard

+
Original3.5Mparameters
Fixture3.5Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

mobilenet_v2.layer

Original · 16
Fixture · 16
KindOriginalFixture
ConvLayer1616✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
mobilenet_v2.layer1.8M51.6%1.8M51.6%
classifier1.3M36.6%1.3M36.6%
mobilenet_v2.conv_1x1412.2K11.8%412.2K11.8%
mobilenet_v2.conv_stem1.8K0.1%1.8K0.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemobilenet_v2mobilenet_v2✓ kept
modelhidden_actrelu6relu6✓ kept
modelimage_size224224✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix MobileNetV2 omitted).

ClassOriginalFixture
BatchNorm2d5252
Conv2d5252
ConvLayer5252
ReLU63535
InvertedResidual1616
AdaptiveAvgPool2d11
Dropout11
Linear11
ForImageClassification11
Model11
Stem11
ModuleList11

Config changes

0 fields changed, 24 unchanged.

FieldOriginalFixture
\ No newline at end of file diff --git a/shrink/mobilenet_v2_1.0_224-mic.json b/shrink/mobilenet_v2_1.0_224-mic.json index 63864f80dbd6082ac847ef4af37dccb24bf4e8b6..4043d973d1f16f4d0aac5c44ad092e4a74d63270 100644 --- a/shrink/mobilenet_v2_1.0_224-mic.json +++ b/shrink/mobilenet_v2_1.0_224-mic.json @@ -1 +1 @@ -{"fixture": "huggingface-mic/mobilenet_v2_1.0_224-mic@91fe07b79191cf1d586a0b7cf0a6e910730a21b6", "findings": ["the fixture keeps 100% of the original's 3.5M parameters"]} \ No newline at end of file +{"fixture": "huggingface-mic/mobilenet_v2_1.0_224-mic@6938bb7a591f10a6640c4ce27183bb9762f016b6", "findings": []} \ No newline at end of file diff --git a/shrink/mobilevit-xx-small-mic.html b/shrink/mobilevit-xx-small-mic.html index b471f140dc4af6e8a74ebd57eca107b5c286f84b..55cd328a20ee7c696d8f85f724523f6b6cb29024 100644 --- a/shrink/mobilevit-xx-small-mic.html +++ b/shrink/mobilevit-xx-small-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

apple/mobilevit-xx-small

huggingface-mic/mobilevit-xx-small-mic@1d4aa4d88fb8 ← apple/mobilevit-xx-small@6703997f9e94 · MobileViTForImageClassification · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 11 failed checks in 30 days · MIC failures dashboard

Original1.3Mparameters
Fixture895.3Kparameters
Layer stacks1compared
Findings1to review

⚠ To review

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

mobilevit.encoder.layer

Original · 5
Fixture · 5
KindOriginalFixture
MobileNetLayer22✓ kept
LayerNorm, Linear, Attention …33✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
mobilevit.encoder.layer924.3K72.7%547.6K61.2%
classifier321.0K25.2%321.0K35.9%
mobilevit.conv_1x1_exp26.2K2.1%26.2K2.9%
mobilevit.conv_stem4640.0%4640.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemobilevitmobilevit✓ kept
modelhidden_actsilusilu✓ kept
modelimage_size256256✓ kept
modelpatch_size22✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix MobileViT omitted).

ClassOriginalFixture
Linear5555
Conv2d3535
ConvLayer3535
SiLUActivation3434
BatchNorm2d3232
Dropout2828
LayerNorm2121
Attention99
Intermediate99
Output99
SelfAttention99
SelfOutput99
TransformerLayer99
InvertedResidual77
ModuleList66
Layer33
Transformer33
MobileNetLayer22
Encoder11
ForImageClassification11
Model11

Config changes

2 fields changed, 28 unchanged.

FieldOriginalFixture
hidden_sizes[64, 80, 96][16, 32, 64]
problem_typenullsingle_label_classification
\ No newline at end of file diff --git a/shrink/modernbert-mic.html b/shrink/modernbert-mic.html index c06254e9d9be5a453c58ea9b8959c00474c34af1..5c876caf048e2d3472a0d86ea3dea7b0707ba954 100644 --- a/shrink/modernbert-mic.html +++ b/shrink/modernbert-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

answerdotai/ModernBERT-base

huggingface-mic/modernbert-mic@59688bef85e5 ← answerdotai/ModernBERT-base@8949b909ec90 · ModernBertForMaskedLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 3 failed checks in 30 days · MIC failures dashboard

Original149.7Mparameters
Fixture3.4Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 22
Fixture · 2
KindOriginalFixture
full_attention81✓ kept
sliding_attention141✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers110.3M73.7%82.1K2.4%
model.embeddings38.7M25.8%3.2M95.9%
head.dense589.8K0.4%4.1K0.1%
decoder50.4K0.0%50.4K1.5%
model.final_norm7680.0%640.0%
head.norm7680.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemodernbertmodernbert✓ kept
modelvocab_size5036850368✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings81928192✓ kept

Module classes

Instances of each class in the built model (prefix ModernBert omitted).

ClassOriginalFixture
Linear9010
LayerNorm466
Dropout233
GELUActivation233
Identity233
Attention222
EncoderLayer222
MLP222
Embedding11
Embeddings11
ForMaskedLM11
Model11
PredictionHead11
RotaryEmbedding11
ModuleList11

Config changes

5 fields changed, 41 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size1152128
layer_types22 entries: 14× sliding_attention, 8× full_attention["full_attention", "sliding_attention"]
num_attention_heads124
num_hidden_layers222
\ No newline at end of file diff --git a/shrink/mpnet-base-mic.html b/shrink/mpnet-base-mic.html index 4fe18574d4a9a6b98c715c8bf169625c9b9d0418..365a734c9661b33faeeda0d796a96476d081abe1 100644 --- a/shrink/mpnet-base-mic.html +++ b/shrink/mpnet-base-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/mpnet-base

huggingface-mic/mpnet-base-mic@9d98c3d72a6c ← microsoft/mpnet-base@6996ce1e91bd · MPNetForMaskedLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original109.5Mparameters
Fixture2.1Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

mpnet.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
mpnet.encoder.layer85.1M77.7%66.9K3.2%
mpnet.embeddings23.8M21.8%2.0M95.1%
lm_head.dense590.6K0.5%4.2K0.2%
lm_head30.5K0.0%30.5K1.5%
lm_head.layer_norm1.5K0.0%1280.0%
mpnet.encoder3840.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typempnetmpnet✓ kept
modelvocab_size3052730527✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings514514✓ kept

Module classes

Instances of each class in the built model (prefix MPNet omitted).

ClassOriginalFixture
Linear7414
Dropout377
LayerNorm266
GELUActivation122
Attention122
Intermediate122
Layer122
Output122
SelfAttention122
Embedding33
Embeddings11
Encoder11
ForMaskedLM11
LMHead11
Model11
ModuleList11

Config changes

4 fields changed, 21 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/mt5-small-mic.html b/shrink/mt5-small-mic.html index b17636cffab8ee5518d8d1b5c716379b85d70338..d563e940efc57e7504be6cad46f808ffb3c46778 100644 --- a/shrink/mt5-small-mic.html +++ b/shrink/mt5-small-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/mt5-small

huggingface-mic/mt5-small-mic@13bc3e6cb33e ← google/mt5-small@73fb5dbe4756 · MT5ForConditionalGeneration · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original172.1Mparameters
Fixture16.2Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

encoder.block

Original · 8
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding71✓ kept

decoder.block

Original · 8
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding71✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
shared128.1M74.4%16.0M98.8%
decoder.block25.2M14.6%115.2K0.7%
encoder.block18.9M11.0%82.3K0.5%
decoder.final_layer_norm5120.0%640.0%
encoder.final_layer_norm5120.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemt5mt5✓ kept
modelvocab_size250112250112✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix MT5 omitted).

ClassOriginalFixture
Linear14537
Dropout5816
LayerNorm4212
Attention246
ModuleList186
Block164
DenseGatedActDense164
LayerFF164
LayerSelfAttention164
NewGELUActivation164
LayerCrossAttention82
Embedding55
Stack22
ForConditionalGeneration11

Config changes

6 fields changed, 27 unchanged.

FieldOriginalFixture
d_ff1024128
d_kv6416
d_model51264
num_decoder_layers82
num_heads64
num_layers82
\ No newline at end of file diff --git a/shrink/nllb-200-distilled-600M-mic.html b/shrink/nllb-200-distilled-600M-mic.html index bb9e624a9099f2c3aea83bba818f1b74124f5456..9e8eca57940587a337a8cc79c12998c9a525131a 100644 --- a/shrink/nllb-200-distilled-600M-mic.html +++ b/shrink/nllb-200-distilled-600M-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/nllb-200-distilled-600M

huggingface-mic/nllb-200-distilled-600M-mic@fcaaa8d89cb3 ← facebook/nllb-200-distilled-600M@f8d333a098d1 · M2M100ForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original615.1Mparameters
Fixture16.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

model.decoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
DecoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.shared262.4M42.7%16.4M99.0%
model.decoder.layers201.6M32.8%100.5K0.6%
model.encoder.layers151.2M24.6%66.9K0.4%
model.encoder2.0K0.0%1280.0%
model.decoder2.0K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typem2m_100m2m_100✓ kept
modelvocab_size256206256206✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings10241024✓ kept

Module classes

Instances of each class in the built model (prefix M2M100 omitted).

ClassOriginalFixture
Linear19333
LayerNorm6212
Attention366
ReLU244
DecoderLayer122
EncoderLayer122
ScaledWordEmbedding33
SinusoidalPositionalEmbedding22
ModuleList22
Decoder11
Encoder11
ForConditionalGeneration11
Model11

Config changes

7 fields changed, 26 unchanged.

FieldOriginalFixture
d_model102464
decoder_attention_heads164
decoder_ffn_dim4096128
decoder_layers122
encoder_attention_heads164
encoder_ffn_dim4096128
encoder_layers122
\ No newline at end of file diff --git a/shrink/nomic-embed-text-v1.5-mic.html b/shrink/nomic-embed-text-v1.5-mic.html index 02a9d04e90e315e082c389fc95089ba9dae59771..436c02257fc86e55baba32b3507b3e73ccb5cbc0 100644 --- a/shrink/nomic-embed-text-v1.5-mic.html +++ b/shrink/nomic-embed-text-v1.5-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

nomic-ai/nomic-embed-text-v1.5

huggingface-mic/nomic-embed-text-v1.5-mic@2cbaf2971908 ← nomic-ai/nomic-embed-text-v1.5@e9b6763023c6 · NomicBertModel · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original136.7Mparameters
Fixture2.0Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

layers

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
layers113.3M82.9%82.4K4.0%
embeddings.word_embeddings23.4M17.1%2.0M95.9%
embeddings.LayerNorm1.5K0.0%1280.0%
embeddings.token_type_embeddings1.5K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelRoPE typedefaultdefault✓ kept
modelmodel_typenomic_bertnomic_bert✓ kept
modelvocab_size3052830528✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings20482048✓ kept

Module classes

Instances of each class in the built model (prefix NomicBert omitted).

ClassOriginalFixture
Linear8414
Dropout255
LayerNorm255
Attention122
Layer122
MLP122
SiLUActivation122
Embedding22
ModuleList11
Embeddings11
Model11
RotaryEmbedding11

Config changes

9 fields changed, 57 unchanged.

FieldOriginalFixture
head_dim6416
hidden_size76864
intermediate_size3072128
n_embd76864
n_head124
n_inner3072128
n_layer122
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/opt-125m-mic.html b/shrink/opt-125m-mic.html index 5d6fc951218b353a0119f0e17972fe7569879a1c..27a86ede5ff9bc72940d6ed68f27ed5c7a5b4531 100644 --- a/shrink/opt-125m-mic.html +++ b/shrink/opt-125m-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/opt-125m

huggingface-mic/opt-125m-mic@30dda77b685d ← facebook/opt-125m@27dcfa74d334 · OPTForCausalLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original125.2Mparameters
Fixture38.9Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.decoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
DecoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.decoder.layers85.1M67.9%66.9K0.2%
model.decoder40.2M32.1%38.8M99.8%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeoptopt✓ kept
modelvocab_size5027250272✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings20482048✓ kept

Module classes

Instances of each class in the built model (prefix OPT omitted).

ClassOriginalFixture
Linear7315
LayerNorm255
Attention122
DecoderLayer122
ReLU122
Embedding11
ModuleList11
Decoder11
ForCausalLM11
LearnedPositionalEmbedding11
Model11

Config changes

4 fields changed, 28 unchanged.

FieldOriginalFixture
ffn_dim3072128
hidden_size76864
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/opus-mt-fr-en-mic.html b/shrink/opus-mt-fr-en-mic.html index 400793b686b33226def38a9bb537236f497a41b9..15a697136edecb7dccafbcedf6f5069c8948f7a3 100644 --- a/shrink/opus-mt-fr-en-mic.html +++ b/shrink/opus-mt-fr-en-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

Helsinki-NLP/opus-mt-fr-en

huggingface-mic/opus-mt-fr-en-mic@d09105cda74d ← Helsinki-NLP/opus-mt-fr-en@c4aed37b318c · MarianMTModel · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original75.1Mparameters
Fixture4.0Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
MarianEncoderLayer62✓ kept

model.decoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
MarianDecoderLayer62✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.shared30.5M40.6%3.8M94.2%
model.decoder.layers25.2M33.6%100.5K2.5%
model.encoder.layers18.9M25.2%66.9K1.7%
model.encoder262.1K0.3%32.8K0.8%
model.decoder262.1K0.3%32.8K0.8%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemarianmarian✓ kept
modelvocab_size5951459514✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings512512✓ kept

Module classes

Instances of each class in the built model (prefix MarianMT omitted).

ClassOriginalFixture
Linear9733
LayerNorm3010
MarianAttention186
SiLU124
MarianDecoderLayer62
MarianEncoderLayer62
Embedding33
MarianSinusoidalPositionalEmbedding22
ModuleList22
MarianDecoder11
MarianEncoder11
Model11
MarianModel11

Config changes

7 fields changed, 38 unchanged.

FieldOriginalFixture
d_model51264
decoder_attention_heads84
decoder_ffn_dim2048128
decoder_layers62
encoder_attention_heads84
encoder_ffn_dim2048128
encoder_layers62
\ No newline at end of file diff --git a/shrink/owlv2-base-patch16-ensemble-mic.html b/shrink/owlv2-base-patch16-ensemble-mic.html index 4c3b72780380d2631d19dec24eac300e19b2b085..8b6d8bbf2d49ec9d77361715a19ae40685f21be7 100644 --- a/shrink/owlv2-base-patch16-ensemble-mic.html +++ b/shrink/owlv2-base-patch16-ensemble-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/owlv2-base-patch16-ensemble

huggingface-mic/owlv2-base-patch16-ensemble-mic@de5f6eb0e6f6 ← google/owlv2-base-patch16-ensemble@cfd3195ba4ea · Owlv2ForObjectDetection · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original155.0Mparameters
Fixture3.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

owlv2.text_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

owlv2.vision_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
owlv2.vision_model.encoder.layers85.1M54.9%66.9K1.9%
owlv2.text_model.encoder.layers37.8M24.4%66.9K1.9%
owlv2.text_model25.3M16.3%3.2M87.7%
owlv2.vision_model3.4M2.2%279.9K7.8%
objectness_head.dense0590.6K0.4%4.2K0.1%
box_head.dense0590.6K0.4%4.2K0.1%
objectness_head.dense1590.6K0.4%4.2K0.1%
10 others1.6M1.1%17.1K0.5%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeowlv2owlv2✓ kept
text_configmodel_typeowlv2_text_modelowlv2_text_model✓ kept
text_configvocab_size4940849408✓ kept
text_confighidden_actquick_geluquick_gelu✓ kept
text_configmax_position_embeddings1616✓ kept
vision_configmodel_typeowlv2_vision_modelowlv2_vision_model✓ kept
vision_confighidden_actquick_geluquick_gelu✓ kept
vision_configimage_size960960✓ kept
vision_configpatch_size1616✓ kept
vision_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Owlv2 omitted).

ClassOriginalFixture
Linear15535
LayerNorm5212
Attention244
EncoderLayer244
MLP244
QuickGELUActivation244
Embedding33
GELU22
ModuleList22
BoxPredictionHead22
Encoder22
Conv2d11
ELU11
ClassPredictionHead11
ForObjectDetection11
Model11
TextEmbeddings11
TextTransformer11
VisionEmbeddings11
VisionTransformer11
Sigmoid11

Config changes

9 fields changed, 47 unchanged.

FieldOriginalFixture
projection_dim51264
text_config.hidden_size51264
text_config.intermediate_size2048128
text_config.num_attention_heads84
text_config.num_hidden_layers122
vision_config.hidden_size76864
vision_config.intermediate_size3072128
vision_config.num_attention_heads124
vision_config.num_hidden_layers122
\ No newline at end of file diff --git a/shrink/owlvit-base-patch32-mic.html b/shrink/owlvit-base-patch32-mic.html index 8019ca6b933ebc7a2df9cf578be02180385b0a07..03e0839c1b72c1bb3319bf51e3ceca08571b6490 100644 --- a/shrink/owlvit-base-patch32-mic.html +++ b/shrink/owlvit-base-patch32-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/owlvit-base-patch32

huggingface-mic/owlvit-base-patch32-mic@09d714a0dd1f ← google/owlvit-base-patch32@cbc355fb3645 · OwlViTForObjectDetection · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original153.2Mparameters
Fixture3.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

owlvit.text_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

owlvit.vision_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
owlvit.vision_model.encoder.layers85.1M55.5%66.9K1.9%
owlvit.text_model.encoder.layers37.8M24.7%66.9K1.9%
owlvit.text_model25.3M16.5%3.2M89.1%
owlvit.vision_model2.8M1.8%233.9K6.6%
box_head.dense1590.6K0.4%4.2K0.1%
box_head.dense0590.6K0.4%4.2K0.1%
class_head.dense0393.7K0.3%4.2K0.1%
7 others661.5K0.4%8.7K0.2%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeowlvitowlvit✓ kept
text_configmodel_typeowlvit_text_modelowlvit_text_model✓ kept
text_configvocab_size4940849408✓ kept
text_configtie_word_embeddingsTrueTrue✓ kept
text_confighidden_actquick_geluquick_gelu✓ kept
text_configmax_position_embeddings1616✓ kept
vision_configmodel_typeowlvit_vision_modelowlvit_vision_model✓ kept
vision_configtie_word_embeddingsTrueTrue✓ kept
vision_confighidden_actquick_geluquick_gelu✓ kept
vision_configimage_size768768✓ kept
vision_configpatch_size3232✓ kept
vision_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix OwlViT omitted).

ClassOriginalFixture
Linear15232
LayerNorm5212
Attention244
EncoderLayer244
MLP244
QuickGELUActivation244
Embedding33
ModuleList22
Encoder22
Conv2d11
ELU11
GELU11
BoxPredictionHead11
ClassPredictionHead11
ForObjectDetection11
Model11
TextEmbeddings11
TextTransformer11
VisionEmbeddings11
VisionTransformer11
Sigmoid11

Config changes

9 fields changed, 84 unchanged.

FieldOriginalFixture
projection_dim51264
text_config.hidden_size51264
text_config.intermediate_size2048128
text_config.num_attention_heads84
text_config.num_hidden_layers122
vision_config.hidden_size76864
vision_config.intermediate_size3072128
vision_config.num_attention_heads124
vision_config.num_hidden_layers122
\ No newline at end of file diff --git a/shrink/paligemma-3b-pt-224-mic.html b/shrink/paligemma-3b-pt-224-mic.html index c5978b403284f57f411eed9020c7edd5176240ba..c22e7a256eb961747b58f6166da0be064be3c63e 100644 --- a/shrink/paligemma-3b-pt-224-mic.html +++ b/shrink/paligemma-3b-pt-224-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

leo009/paligemma-3b-pt-224

huggingface-mic/paligemma-3b-pt-224-mic@adfd7e09d295 ← leo009/paligemma-3b-pt-224@39996beb6fb1 · PaliGemmaForConditionalGeneration · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original2.9Bparameters
Fixture16.7Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.vision_tower.encoder.layers

Original · 27
Fixture · 2
KindOriginalFixture
SiglipEncoderLayer272✓ kept

model.language_model.layers

Original · 18
Fixture · 2
KindOriginalFixture
GemmaDecoderLayer182✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.language_model.layers2.0B67.8%69.9K0.4%
model.language_model526.8M18.0%16.5M98.8%
model.vision_tower.encoder.layers411.5M14.1%66.9K0.4%
model.multi_modal_projector2.4M0.1%4.2K0.0%
model.vision_tower975.7K0.0%54.2K0.3%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typepaligemmapaligemma✓ kept
modelvocab_size257216257216✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
vision_configmodel_typesiglip_vision_modelsiglip_vision_model✓ kept
vision_confighidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
vision_configimage_size224224✓ kept
vision_configpatch_size1414✓ kept
vision_confignum_channels33✓ kept
text_configattention styleMQA (8 heads, 1 KV)MQA (4 heads, 1 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typegemmagemma✓ kept
text_configvocab_size257216257216✓ kept
text_configtie_word_embeddingsTrueTrue✓ kept
text_confighidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
text_configmax_position_embeddings81928192✓ kept

Module classes

Instances of each class in the built model (prefix PaliGemma omitted).

ClassOriginalFixture
Linear29028
LayerNorm555
GELUTanh454
GemmaRMSNorm375
SiglipAttention272
SiglipEncoderLayer272
SiglipMLP272
GemmaAttention182
GemmaDecoderLayer182
GemmaMLP182
ModuleList22
Conv2d11
Embedding11
GemmaModel11
GemmaRotaryEmbedding11
GemmaTextScaledWordEmbedding11
ForConditionalGeneration11
Model11
MultiModalProjector11
SiglipEncoder11
SiglipVisionEmbeddings11
SiglipVisionModel11

Config changes

12 fields changed, 59 unchanged.

FieldOriginalFixture
hidden_size204864
projection_dim204864
text_config.head_dim25616
text_config.hidden_size204864
text_config.intermediate_size16384128
text_config.num_attention_heads84
text_config.num_hidden_layers182
vision_config.hidden_size115264
vision_config.intermediate_size4304128
vision_config.num_attention_heads164
vision_config.num_hidden_layers272
vision_config.projection_dim204864
\ No newline at end of file diff --git a/shrink/pegasus-xsum-mic.html b/shrink/pegasus-xsum-mic.html index 93f859fb920bb404411d9d9f800597d8eec85252..1a5b8b779686f73018e36c25c7c10b5edb06f2dc 100644 --- a/shrink/pegasus-xsum-mic.html +++ b/shrink/pegasus-xsum-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/pegasus-xsum

huggingface-mic/pegasus-xsum-mic@3703172718d0 ← google/pegasus-xsum@8d8ffc158a3b · PegasusForConditionalGeneration · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original569.7Mparameters
Fixture6.4Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 16
Fixture · 2
KindOriginalFixture
EncoderLayer162✓ kept

model.decoder.layers

Original · 16
Fixture · 2
KindOriginalFixture
DecoderLayer162✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.decoder.layers268.7M47.2%100.5K1.6%
model.encoder.layers201.5M35.4%66.9K1.0%
model.shared98.4M17.3%6.2M96.3%
model.decoder526.3K0.1%32.9K0.5%
model.encoder526.3K0.1%32.9K0.5%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typepegasuspegasus✓ kept
modelvocab_size9610396103✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings512512✓ kept

Module classes

Instances of each class in the built model (prefix Pegasus omitted).

ClassOriginalFixture
Linear25733
LayerNorm8212
Attention486
ReLU324
DecoderLayer162
EncoderLayer162
Embedding33
ModuleList22
SinusoidalPositionalEmbedding22
Decoder11
Encoder11
ForConditionalGeneration11
Model11

Config changes

7 fields changed, 38 unchanged.

FieldOriginalFixture
d_model102464
decoder_attention_heads164
decoder_ffn_dim4096128
decoder_layers162
encoder_attention_heads164
encoder_ffn_dim4096128
encoder_layers162
\ No newline at end of file diff --git a/shrink/phi-2-mic.html b/shrink/phi-2-mic.html index 588cc456ece014f809baefe90b62e416f0c515f4..f776d9c050f13834d7c0db0b07b5a1cc04f93ed8 100644 --- a/shrink/phi-2-mic.html +++ b/shrink/phi-2-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/phi-2

huggingface-mic/phi-2-mic@af26f3c429a4 ← microsoft/phi-2@810d367871c1 · PhiForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original2.8Bparameters
Fixture6.7Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 32
Fixture · 2
KindOriginalFixture
DecoderLayer322✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers2.5B90.6%66.7K1.0%
lm_head131.1M4.7%3.3M49.9%
model.embed_tokens131.1M4.7%3.3M49.1%
model.final_layernorm5.1K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleMHA (32 heads, 32 KV)MHA (4 heads, 4 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typephiphi✓ kept
modelvocab_size5120051200✓ kept
modeltie_word_embeddingsFalseFalse✓ kept
modelhidden_actgelu_newgelu_new✓ kept
modelmax_position_embeddings20482048✓ kept

Module classes

Instances of each class in the built model (prefix Phi omitted).

ClassOriginalFixture
Linear19313
Dropout333
LayerNorm333
NewGELUActivation322
Attention322
DecoderLayer322
MLP322
Embedding11
ModuleList11
ForCausalLM11
Model11
RotaryEmbedding11

Config changes

5 fields changed, 24 unchanged.

FieldOriginalFixture
hidden_size256064
intermediate_size10240128
num_attention_heads324
num_hidden_layers322
num_key_value_heads324
\ No newline at end of file diff --git a/shrink/privacy-filter-mic.html b/shrink/privacy-filter-mic.html index bdf98455a48d08f91e98caf8359d3b98c3e03979..256f573dcf1cadcfba1ae7b74c4dfa510a6c2397 100644 --- a/shrink/privacy-filter-mic.html +++ b/shrink/privacy-filter-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

openai/privacy-filter

huggingface-mic/privacy-filter-mic@ad3dba8d001d ← openai/privacy-filter@7ffa9a043d54 · OpenAIPrivacyFilterForTokenClassification · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original1.4Bparameters
Fixture13.0Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 8
Fixture · 2
KindOriginalFixture
EncoderLayer82✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers1.3B90.8%224.9K1.7%
model.embed_tokens128.0M9.1%12.8M98.3%
score21.2K0.0%2.1K0.0%
model.norm6400.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleGQA (14 heads, 2 KV)GQA (4 heads, 2 KV)✓ kept
modelMoE routingtop-4 of 128 expertstop-2 of 4 experts✓ kept
modelRoPE typeyarnyarn✓ kept
modelmodel_typeopenai_privacy_filteropenai_privacy_filter✓ kept
modelvocab_size200064200064✓ kept
modeltie_word_embeddingsFalseFalse✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings131072131072✓ kept

Module classes

Instances of each class in the built model (prefix OpenAIPrivacyFilter omitted).

ClassOriginalFixture
Linear339
RMSNorm175
Attention82
EncoderLayer82
Experts82
MLP82
TopKRouter82
Dropout11
Embedding11
ModuleList11
ForTokenClassification11
Model11
RotaryEmbedding11

Config changes

7 fields changed, 32 unchanged.

FieldOriginalFixture
head_dim6416
hidden_size64064
intermediate_size640128
num_attention_heads144
num_experts_per_tok42
num_hidden_layers82
num_local_experts1284
\ No newline at end of file diff --git a/shrink/pvt-tiny-224-mic.html b/shrink/pvt-tiny-224-mic.html index 8e34ea11bf239ac120cd34ddc1961abdacdd735d..874ed50dd7ee6997f00d742ae95615a8f8923d76 100644 --- a/shrink/pvt-tiny-224-mic.html +++ b/shrink/pvt-tiny-224-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

Zetatech/pvt-tiny-224

huggingface-mic/pvt-tiny-224-mic@b1c239c03ea2 ← Zetatech/pvt-tiny-224@6a44ba96dda6 · PvtForImageClassification · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original13.2Mparameters
Fixture592.2Kparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

pvt.encoder.patch_embeddings

Original · 4
Fixture · 4
KindOriginalFixture
Conv2d44✓ kept

pvt.encoder.block

Original · 4
Fixture · 4
KindOriginalFixture
Conv2d33✓ kept
no Conv2d11✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
pvt.encoder.block11.5M86.7%324.2K54.7%
pvt.encoder.patch_embeddings1.2M9.4%138.8K23.4%
classifier513.0K3.9%129.0K21.8%
pvt.encoder1.0K0.0%2560.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typepvtpvt✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size224224✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Pvt omitted).

ClassOriginalFixture
Linear4925
Dropout2816
LayerNorm2716
Conv2d107
GELUActivation84
Identity84
Attention84
EfficientSelfAttention84
FFN84
Layer84
SelfOutput84
ModuleList66
PatchEmbeddings44
Encoder11
ForImageClassification11
Model11

Config changes

4 fields changed, 24 unchanged.

FieldOriginalFixture
depths[2, 2, 2, 2][1, 1, 1, 1]
hidden_sizes[64, 128, 320, 512][16, 32, 64, 128]
num_attention_heads[1, 2, 5, 8][1, 2, 4, 8]
problem_typenullsingle_label_classification
\ No newline at end of file diff --git a/shrink/pythia-160m-mic.html b/shrink/pythia-160m-mic.html index bd563568a55aae31908597d2c878f8dede0da504..10303cdf8a90c2e8b47ec53e79d2c62ff28b52c2 100644 --- a/shrink/pythia-160m-mic.html +++ b/shrink/pythia-160m-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

EleutherAI/pythia-160m

huggingface-mic/pythia-160m-mic@f21c9ac8c4e8 ← EleutherAI/pythia-160m@50f5173d932e · GPTNeoXForCausalLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original162.3Mparameters
Fixture6.5Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

gpt_neox.layers

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
gpt_neox.layers85.1M52.4%66.9K1.0%
gpt_neox.embed_in38.6M23.8%3.2M49.5%
lm_head38.6M23.8%3.2M49.5%
gpt_neox.final_layer_norm1.5K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelRoPE typedefaultdefault✓ kept
modelmodel_typegpt_neoxgpt_neox✓ kept
modelvocab_size5030450304✓ kept
modeltie_word_embeddingsFalseFalse✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings20482048✓ kept

Module classes

Instances of each class in the built model (prefix GPTNeoX omitted).

ClassOriginalFixture
Linear499
Dropout255
LayerNorm255
GELUActivation122
Attention122
Layer122
MLP122
Embedding11
ForCausalLM11
Model11
RotaryEmbedding11
ModuleList11

Config changes

4 fields changed, 26 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/qwen2-mic.html b/shrink/qwen2-mic.html index e7105ec28ca8b7f2005731d4afe8c98922ce1c87..fe23df30788edff86dcf4db9f1c703c1c06fcbd3 100644 --- a/shrink/qwen2-mic.html +++ b/shrink/qwen2-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

Qwen/Qwen2-0.5B

huggingface-mic/qwen2-mic@f9c74f62c625 ← Qwen/Qwen2-0.5B@91d2aff3f957 · Qwen2ForCausalLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 2 failed checks in 30 days · MIC failures dashboard

Original494.0Mparameters
Fixture9.8Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 24
Fixture · 2
KindOriginalFixture
DecoderLayer242✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers357.9M72.4%74.2K0.8%
model.embed_tokens136.1M27.6%9.7M99.2%
model.norm8960.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleGQA (14 heads, 2 KV)GQA (4 heads, 2 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typeqwen2qwen2✓ kept
modelvocab_size151936151936✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings131072131072✓ kept

Module classes

Instances of each class in the built model (prefix Qwen2 omitted).

ClassOriginalFixture
Linear16915
RMSNorm495
Attention242
DecoderLayer242
MLP242
SiLUActivation242
Embedding11
ModuleList11
ForCausalLM11
Model11
RotaryEmbedding11

Config changes

5 fields changed, 25 unchanged.

FieldOriginalFixture
hidden_size89664
intermediate_size4864128
layer_types24 entries: 24× full_attention["full_attention", "full_attention"]
num_attention_heads144
num_hidden_layers242
\ No newline at end of file diff --git a/shrink/qwen2_5_omni-mic.html b/shrink/qwen2_5_omni-mic.html index 02230a49077fbab48b1d3b03d8c8a912c5f10a1e..53ff3672f03ff8bfd0f3c2b90f03347a437c7d42 100644 --- a/shrink/qwen2_5_omni-mic.html +++ b/shrink/qwen2_5_omni-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

Qwen/Qwen2.5-Omni-7B

huggingface-mic/qwen2_5_omni-mic@72e87592273c ← Qwen/Qwen2.5-Omni-7B@ae9e1690543f · Qwen2_5OmniForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: failing now on main (26 failed checks in 30 days) · MIC failures dashboard

Original10.7Bparameters
Fixture170.4Mparameters
Layer stacks8compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

thinker.model.audio_tower.layers

Original · 32
Fixture · 2
KindOriginalFixture
AudioEncoderLayer322✓ kept

thinker.model.visual.blocks

Original · 32
Fixture · 32
KindOriginalFixture
VisionBlock3232✓ kept

thinker.model.language_model.layers

Original · 28
Fixture · 2
KindOriginalFixture
DecoderLayer282✓ kept

talker.model.layers

Original · 24
Fixture · 2
KindOriginalFixture
DecoderLayer242✓ kept

token2wav.code2wav_dit_model.input_embed.spk_encoder.blocks

Original · 4
Fixture · 4
KindOriginalFixture
no ModuleList11✓ kept
ModuleList, Res2NetBlock, Sigmoid …33✓ kept

token2wav.code2wav_dit_model.transformer_blocks

Original · 22
Fixture · 2
KindOriginalFixture
DiTDecoderLayer222✓ kept

token2wav.code2wav_bigvgan_model.ups

Original · 6
Fixture · 6
KindOriginalFixture
ModuleList66✓ kept

token2wav.code2wav_bigvgan_model.resblocks

Original · 18
Fixture · 18
KindOriginalFixture
AMPBlock1818✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
thinker.model.language_model.layers6.5B60.8%74.2K0.0%
talker.model.layers1.3B12.2%74.2K0.0%
thinker.model.visual.blocks630.5M5.9%1.3M0.8%
thinker.model.audio_tower.layers629.6M5.9%66.8K0.0%
thinker.model601.1M5.6%10.2M6.0%
thinker.lm_head545.0M5.1%9.7M5.7%
token2wav.code2wav_dit_model.transformer_blocks323.3M3.0%116.4K0.1%
8 others166.9M1.6%148.8M87.3%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeqwen2_5_omniqwen2_5_omni✓ kept
thinker_configmodel_typeqwen2_5_omni_thinkerqwen2_5_omni_thinker✓ kept
thinker_configtie_word_embeddingsFalseFalse✓ kept
thinker_config.audio_configmodel_typeqwen2_5_omni_audio_encoderqwen2_5_omni_audio_encoder✓ kept
thinker_config.audio_configtie_word_embeddingsTrueTrue✓ kept
thinker_config.audio_confignum_mel_bins128128✓ kept
thinker_config.vision_configRoPE typeaxialaxial✓ kept
thinker_config.vision_configmodel_typeqwen2_5_omni_vision_encoderqwen2_5_omni_vision_encoder✓ kept
thinker_config.vision_configtie_word_embeddingsTrueTrue✓ kept
thinker_config.vision_confighidden_actsilusilu✓ kept
thinker_config.vision_configpatch_size1414✓ kept
thinker_config.text_configattention styleGQA (28 heads, 4 KV)GQA (4 heads, 2 KV)✓ kept
thinker_config.text_configRoPE typedefaultdefault✓ kept
thinker_config.text_configmodel_typeqwen2_5_omni_textqwen2_5_omni_text✓ kept
thinker_config.text_configvocab_size152064152064✓ kept
thinker_config.text_configtie_word_embeddingsFalseFalse✓ kept
thinker_config.text_confighidden_actsilusilu✓ kept
thinker_config.text_configmax_position_embeddings3276832768✓ kept
talker_configattention styleGQA (12 heads, 4 KV)GQA (4 heads, 2 KV)✓ kept
talker_configRoPE typedefaultdefault✓ kept
talker_configmodel_typeqwen2_5_omni_talkerqwen2_5_omni_talker✓ kept
talker_configvocab_size84488448✓ kept
talker_configtie_word_embeddingsFalseFalse✓ kept
talker_confighidden_actsilusilu✓ kept
talker_configmax_position_embeddings3276832768✓ kept
token2wav_configmodel_typeqwen2_5_omni_token2wavqwen2_5_omni_token2wav✓ kept
token2wav_config.dit_configRoPE typedefaultdefault✓ kept
token2wav_config.dit_configmodel_typeqwen2_5_omni_ditqwen2_5_omni_dit✓ kept
token2wav_config.dit_configtie_word_embeddingsTrueTrue✓ kept
token2wav_config.dit_configmax_position_embeddings3276832768✓ kept
token2wav_config.bigvgan_configmodel_typeqwen2_5_omni_bigvganqwen2_5_omni_bigvgan✓ kept
token2wav_config.bigvgan_configtie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix Qwen2_5Omni omitted).

ClassOriginalFixture
Linear945289
RMSNorm17175
Conv1d132132
ModuleList11676
LayerNorm11010
AntiAliasedActivation1d109109
DownSample1d109109
SnakeBeta109109
UpSample1d109109
SiLUActivation8436
Qwen2MLP524
Attention524
DecoderLayer524
Dropout444
GELUActivation322
AudioAttention322
AudioEncoderLayer322
MLP3232
VisionAttention3232
VisionBlock3232
SiLU244
GELU233
DiTAttention222
DiTDecoderLayer222
DiTMLP222
Qwen2_5_OmniAdaLayerNormZero222
AMPBlock1818
ReLU1515
TimeDelayNetBlock1212
ConvTranspose1d66
Embedding44
Res2NetBlock33
Sigmoid33
SqueezeExcitationBlock33
SqueezeExcitationRes2NetBlock33
RotaryEmbedding22
AttentiveStatisticsPooling11
AvgPool1d11
Conv3d11
DiTCodecEmbedding11
DiTInputEmbedding11
DiTTimestepEmbedding11
ECAPA_TimeDelayNet11
AudioEncoder11
DiTRotaryEmbedding11
ForConditionalGeneration11
PatchMerger11
TalkerForConditionalGeneration11
TalkerModel11
ThinkerForConditionalGeneration11
ThinkerModel11
ThinkerTextModel11
Token2WavBigVGANModel11
Token2WavDiTModel11
Token2WavModel11
VisionEncoder11
VisionRotaryEmbedding11
Qwen2_5_OmniAdaLayerNormZero_Final11
Qwen2_5_VisionPatchEmbed11
Sequential11
SinusPositionEmbedding11
SinusoidsPositionEmbedding11
Tanh11

Config changes

30 fields changed, 281 unchanged.

FieldOriginalFixture
talker_config.head_dim12816
talker_config.hidden_size89664
talker_config.intermediate_size18944128
talker_config.layer_types24 entries: 24× full_attention["full_attention", "full_attention"]
talker_config.num_attention_heads124
talker_config.num_hidden_layers242
talker_config.num_key_value_heads42
talker_config.rope_parameters{"mrope_section": [16, 24, 24], "rope_type": "default", "type": "default", "r…{"mrope_section": [2, 3, 3], "rope_theta": 1000000.0, "rope_type": "default",…
thinker_config.audio_config.d_model128064
thinker_config.audio_config.encoder_attention_heads204
thinker_config.audio_config.encoder_ffn_dim5120128
thinker_config.audio_config.encoder_layers322
thinker_config.text_config.hidden_size358464
thinker_config.text_config.intermediate_size18944128
thinker_config.text_config.layer_types28 entries: 28× full_attention["full_attention", "full_attention"]
thinker_config.text_config.num_attention_heads284
thinker_config.text_config.num_hidden_layers282
thinker_config.text_config.num_key_value_heads42
thinker_config.text_config.rope_parameters{"mrope_section": [16, 24, 24], "rope_type": "default", "type": "default", "r…{"mrope_section": [2, 3, 3], "rope_theta": 1000000.0, "rope_type": "default",…
thinker_config.vision_config.embed_dim128064
thinker_config.vision_config.hidden_size128064
thinker_config.vision_config.intermediate_size3420128
thinker_config.vision_config.num_heads164
thinker_config.vision_config.out_hidden_size358464
token2wav_config.dit_config.dim102464
token2wav_config.dit_config.emb_dim51264
token2wav_config.dit_config.head_dim6416
token2wav_config.dit_config.hidden_size102464
token2wav_config.dit_config.num_attention_heads164
token2wav_config.dit_config.num_hidden_layers222
\ No newline at end of file diff --git a/shrink/qwen2_5_vl-mic.html b/shrink/qwen2_5_vl-mic.html index 721d93ede993b6b49a70324f1061a3e2b6109a9e..22dca89d99a7fe6473f0705db4159275f6ebee0a 100644 --- a/shrink/qwen2_5_vl-mic.html +++ b/shrink/qwen2_5_vl-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

Qwen/Qwen2.5-VL-3B-Instruct

huggingface-mic/qwen2_5_vl-mic@32d6627e4ea6 ← Qwen/Qwen2.5-VL-3B-Instruct@66285546d2b8 · Qwen2_5_VLForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 2 failed checks in 30 days · MIC failures dashboard

Original3.8Bparameters
Fixture11.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.visual.blocks

Original · 32
Fixture · 32
KindOriginalFixture
VisionBlock3232✓ kept

model.language_model.layers

Original · 36
Fixture · 2
KindOriginalFixture
DecoderLayer362✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.language_model.layers2.8B73.9%74.2K0.7%
model.visual.blocks630.5M16.8%1.3M11.8%
model.language_model311.2M8.3%9.7M86.1%
model.visual38.2M1.0%157.6K1.4%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeqwen2_5_vlqwen2_5_vl✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
text_configattention styleGQA (16 heads, 2 KV)GQA (4 heads, 2 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typeqwen2_5_vl_textqwen2_5_vl_text✓ kept
text_configvocab_size151936151936✓ kept
text_confighidden_actsilusilu✓ kept
text_configmax_position_embeddings128000128000✓ kept
vision_configRoPE typeaxialaxial✓ kept
vision_configmodel_typeqwen2_5_vl_visionqwen2_5_vl_vision✓ kept
vision_confighidden_actsilusilu✓ kept
vision_configpatch_size1414✓ kept

Module classes

Instances of each class in the built model (prefix Qwen2_5_VL omitted).

ClassOriginalFixture
Linear415177
RMSNorm13870
SiLUActivation6834
Qwen2MLP362
Attention362
DecoderLayer362
MLP3232
VisionAttention3232
VisionBlock3232
ModuleList22
Conv3d11
Embedding11
GELU11
ForConditionalGeneration11
Model11
PatchMerger11
RotaryEmbedding11
TextModel11
VisionRotaryEmbedding11
Qwen2_5_VisionPatchEmbed11
Qwen2_5_VisionTransformerPretrainedModel11
Sequential11

Config changes

10 fields changed, 60 unchanged.

FieldOriginalFixture
text_config.hidden_size204864
text_config.intermediate_size11008128
text_config.layer_types36 entries: 36× full_attention["full_attention", "full_attention"]
text_config.num_attention_heads164
text_config.num_hidden_layers362
text_config.rope_parameters{"type": "mrope", "mrope_section": [16, 24, 24], "rope_theta": 1000000.0, "ro…{"mrope_section": [2, 3, 3], "rope_theta": 1000000.0, "rope_type": "default",…
vision_config.hidden_size128064
vision_config.intermediate_size3420128
vision_config.num_heads164
vision_config.out_hidden_size204864
\ No newline at end of file diff --git a/shrink/qwen2_audio-mic.html b/shrink/qwen2_audio-mic.html index fcbfbe0bf76177c0527957c3cf1ddd0c956c890f..7adc69f85196f51f43850450aec54e294ff521ef 100644 --- a/shrink/qwen2_audio-mic.html +++ b/shrink/qwen2_audio-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

Qwen/Qwen2-Audio-7B

huggingface-mic/qwen2_audio-mic@c426549756a1 ← Qwen/Qwen2-Audio-7B@dd84470756e6 · Qwen2AudioForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 3 failed checks in 30 days · MIC failures dashboard

Original8.4Bparameters
Fixture20.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.audio_tower.layers

Original · 32
Fixture · 2
KindOriginalFixture
EncoderLayer322✓ kept

model.language_model.layers

Original · 32
Fixture · 2
KindOriginalFixture
Qwen2DecoderLayer322✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.language_model.layers6.5B77.1%82.6K0.4%
model.language_model639.1M7.6%10.0M49.3%
lm_head639.1M7.6%10.0M49.3%
model.audio_tower.layers629.6M7.5%66.8K0.3%
model.audio_tower7.3M0.1%133.1K0.7%
model.multi_modal_projector5.2M0.1%4.2K0.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeqwen2_audioqwen2_audio✓ kept
modelvocab_size156032156032✓ kept
audio_configmodel_typeqwen2_audio_encoderqwen2_audio_encoder✓ kept
audio_confignum_mel_bins128128✓ kept
text_configattention styleMHA (32 heads, 32 KV)MHA (4 heads, 4 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typeqwen2qwen2✓ kept
text_configvocab_size156032156032✓ kept
text_configtie_word_embeddingsFalseFalse✓ kept
text_confighidden_actsilusilu✓ kept
text_configmax_position_embeddings81928192✓ kept

Module classes

Instances of each class in the built model (prefix Qwen2Audio omitted).

ClassOriginalFixture
Linear41828
LayerNorm655
Qwen2RMSNorm655
GELUActivation322
Qwen2Attention322
Attention322
EncoderLayer322
Qwen2DecoderLayer322
Qwen2MLP322
SiLUActivation322
Conv1d22
Embedding22
ModuleList22
AvgPool1d11
Encoder11
ForConditionalGeneration11
Model11
MultiModalProjector11
Qwen2Model11
Qwen2RotaryEmbedding11

Config changes

10 fields changed, 55 unchanged.

FieldOriginalFixture
audio_config.d_model128064
audio_config.encoder_attention_heads204
audio_config.encoder_ffn_dim5120128
audio_config.encoder_layers322
text_config.hidden_size409664
text_config.intermediate_size11008128
text_config.layer_types32 entries: 32× full_attention["full_attention", "full_attention"]
text_config.num_attention_heads324
text_config.num_hidden_layers322
text_config.num_key_value_heads324
\ No newline at end of file diff --git a/shrink/recurrentgemma-2b-mic.html b/shrink/recurrentgemma-2b-mic.html index d50a23a07f70efcaf637dc09160cde667b521875..740aa79d402ee95be31e55b0827d0e70ffa8d7c4 100644 --- a/shrink/recurrentgemma-2b-mic.html +++ b/shrink/recurrentgemma-2b-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/recurrentgemma-2b

huggingface-mic/recurrentgemma-2b-mic@732c0b9399dc ← google/recurrentgemma-2b@3620f4ca9c5d · RecurrentGemmaForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original2.7Bparameters
Fixture24.0Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 26
Fixture · 3
KindOriginalFixture
Conv1d, RecurrentBlock, Rglru182✓ kept
is_causal=True, Attention, RotaryEmbedding81✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers2.0B75.6%7.6M31.8%
model.embed_tokens655.4M24.4%16.4M68.2%
model.final_norm2.6K0.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleMQA (10 heads, 1 KV)MQA (4 heads, 1 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typerecurrent_gemmarecurrent_gemma✓ kept
modelvocab_size256000256000✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix RecurrentGemma omitted).

ClassOriginalFixture
Linear16520
RMSNorm537
GELUTanh445
DecoderLayer263
Mlp263
Conv1d182
RecurrentBlock182
Rglru182
Attention81
RotaryEmbedding81
Embedding11
ModuleList11
ForCausalLM11
Model11

Config changes

5 fields changed, 30 unchanged.

FieldOriginalFixture
head_dim25616
hidden_size256064
intermediate_size15360128
num_attention_heads104
num_hidden_layers263
\ No newline at end of file diff --git a/shrink/resnet-18-mic.html b/shrink/resnet-18-mic.html index 0560bbfb02cab798f3f941a09c6ddbec6c83b9b3..329e9883304457b3e57629245ce29fe7623e981e 100644 --- a/shrink/resnet-18-mic.html +++ b/shrink/resnet-18-mic.html @@ -31,6 +31,7 @@ tr.bad td{color:var(--warn)} .c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}

← MIC coverage of Transformers model types

-

microsoft/resnet-18

huggingface-mic/resnet-18-mic@31e1f7325db3 ← microsoft/resnet-18@65a5785d9156 · ResNetForImageClassification · overfit weights

-
Original11.7Mparameters
Fixture438.5Kparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

resnet.encoder.stages

Original · 4
Fixture · 4
KindOriginalFixture
no ShortCut11✓ kept
ShortCut33✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
resnet.encoder.stages11.2M95.5%307.1K70.0%
classifier.1513.0K4.4%129.0K29.4%
resnet.embedder9.5K0.1%2.4K0.5%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeresnetresnet✓ kept
modelhidden_actrelurelu✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix ResNet omitted).

ClassOriginalFixture
BatchNorm2d2012
Conv2d2012
ReLU179
ConvLayer179
Identity135
Sequential139
BasicLayer84
Stage44
ShortCut33
AdaptiveAvgPool2d11
Flatten11
Linear11
MaxPool2d11
ModuleList11
Embeddings11
Encoder11
ForImageClassification11
Model11

Config changes

4 fields changed, 16 unchanged.

FieldOriginalFixture
depths[2, 2, 2, 2][1, 1, 1, 1]
embedding_size6416
hidden_sizes[64, 128, 256, 512][16, 32, 64, 128]
problem_typenullsingle_label_classification
+

microsoft/resnet-18

huggingface-mic/resnet-18-mic@79f172823370 ← microsoft/resnet-18@65a5785d9156 · ResNetForImageClassification · real weights

+

CI: no failure on main in 30 days · MIC failures dashboard

+
Original11.7Mparameters
Fixture11.7Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

resnet.encoder.stages

Original · 4
Fixture · 4
KindOriginalFixture
no ShortCut11✓ kept
ShortCut33✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
resnet.encoder.stages11.2M95.5%11.2M95.5%
classifier.1513.0K4.4%513.0K4.4%
resnet.embedder9.5K0.1%9.5K0.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeresnetresnet✓ kept
modelhidden_actrelurelu✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix ResNet omitted).

ClassOriginalFixture
BatchNorm2d2020
Conv2d2020
ReLU1717
ConvLayer1717
Identity1313
Sequential1313
BasicLayer88
Stage44
ShortCut33
AdaptiveAvgPool2d11
Flatten11
Linear11
MaxPool2d11
ModuleList11
Embeddings11
Encoder11
ForImageClassification11
Model11

Config changes

0 fields changed, 20 unchanged.

FieldOriginalFixture
\ No newline at end of file diff --git a/shrink/resnet-18-mic.json b/shrink/resnet-18-mic.json index 0f45d0a99d2c98f523a5cffefbe75348cf43c572..80fcd00d8d8705c92629d9dbdb80f0bc5cc76497 100644 --- a/shrink/resnet-18-mic.json +++ b/shrink/resnet-18-mic.json @@ -1 +1 @@ -{"fixture": "huggingface-mic/resnet-18-mic@31e1f7325db335a5c638bb94f40f84dd9977277b", "findings": []} \ No newline at end of file +{"fixture": "huggingface-mic/resnet-18-mic@79f17282337052447d8e69c74df99f801e596790", "findings": []} \ No newline at end of file diff --git a/shrink/roberta-base-mic.html b/shrink/roberta-base-mic.html index 64b3a53545db294fdef4152333623def5a8c80e7..4d02709a9da334d864db588dc9ee44fe132d3e10 100644 --- a/shrink/roberta-base-mic.html +++ b/shrink/roberta-base-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

FacebookAI/roberta-base

huggingface-mic/roberta-base-mic@ff3c8d1292ae ← FacebookAI/roberta-base@e2da8e2f811d · RobertaForMaskedLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original124.7Mparameters
Fixture3.4Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

roberta.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
roberta.encoder.layer85.1M68.2%66.9K2.0%
roberta.embeddings39.0M31.3%3.3M96.4%
lm_head.dense590.6K0.5%4.2K0.1%
lm_head50.3K0.0%50.3K1.5%
lm_head.layer_norm1.5K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typerobertaroberta✓ kept
modelvocab_size5026550265✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings514514✓ kept

Module classes

Instances of each class in the built model (prefix Roberta omitted).

ClassOriginalFixture
Linear7414
Dropout377
LayerNorm266
GELUActivation122
Attention122
Intermediate122
Layer122
Output122
SelfAttention122
SelfOutput122
Embedding33
ModuleList11
Embeddings11
Encoder11
ForMaskedLM11
LMHead11
Model11

Config changes

4 fields changed, 25 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/rtdetr_r50vd-mic.html b/shrink/rtdetr_r50vd-mic.html index cfffb2e26bd237780215cec81c37b613130756ba..883af842e2a2144c7e43a6001f14b2126fbaf95d 100644 --- a/shrink/rtdetr_r50vd-mic.html +++ b/shrink/rtdetr_r50vd-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

PekingU/rtdetr_r50vd

huggingface-mic/rtdetr_r50vd-mic@5655fbe502f7 ← PekingU/rtdetr_r50vd@df939e661d8c · RTDetrForObjectDetection · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original42.9Mparameters
Fixture10.9Mparameters
Layer stacks10compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.backbone.model.encoder.stages

Original · 4
Fixture · 4
KindOriginalFixture
no AvgPool2d11✓ kept
AvgPool2d33✓ kept

model.encoder_input_proj

Original · 3
Fixture · 3
KindOriginalFixture
Conv2d33✓ kept

model.encoder.aifi

Original · 1
Fixture · 1
KindOriginalFixture
AIFILayer11✓ kept

model.encoder.lateral_convs

Original · 2
Fixture · 2
KindOriginalFixture
ConvNormLayer22✓ kept

model.encoder.fpn_blocks

Original · 2
Fixture · 2
KindOriginalFixture
RepVggBlock22✓ kept

model.encoder.downsample_convs

Original · 2
Fixture · 2
KindOriginalFixture
ConvNormLayer22✓ kept

model.encoder.pan_blocks

Original · 2
Fixture · 2
KindOriginalFixture
RepVggBlock22✓ kept

model.decoder_input_proj

Original · 3
Fixture · 3
KindOriginalFixture
Conv2d33✓ kept

model.decoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
DecoderLayer62✓ kept

model.decoder.bbox_embed

Original · 6
Fixture · 2
KindOriginalFixture
Linear62✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.backbone.model.encoder.stages23.4M54.7%32.4K0.3%
model.decoder.layers6.0M13.9%102.6K0.9%
model.encoder.pan_blocks4.5M10.4%4.5M40.9%
model.encoder.fpn_blocks4.5M10.4%4.5M40.9%
model.encoder.downsample_convs1.2M2.8%1.2M10.8%
model.encoder_input_proj919.0K2.1%58.9K0.5%
model.decoder.bbox_embed795.7K1.9%17.2K0.2%
9 others1.6M3.8%582.8K5.3%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typert_detrrt_detr✓ kept
backbone_configmodel_typert_detr_resnetrt_detr_resnet✓ kept
backbone_confighidden_actrelurelu✓ kept
backbone_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix RTDetr omitted).

ClassOriginalFixture
Conv2d9761
Linear9741
ReLU5717
Identity5632
FrozenBatchNorm2d5519
ResNetConvLayer5115
BatchNorm2d4242
ConvNormLayer3636
Sequential3523
SiLUActivation2424
LayerNorm219
ModuleList2016
ResNetBottleNeckLayer164
RepVggBlock1212
MLPPredictionHead84
MLP73
SelfAttention73
MultiScaleDeformableAttention62
DecoderLayer62
MultiscaleDeformableAttention62
CSPRepLayer44
ResNetShortCut44
ResNetStage44
AvgPool2d33
Embedding11
GELUActivation11
MaxPool2d11
AIFILayer11
ConvEncoder11
Decoder11
EncoderLayer11
ForObjectDetection11
HybridEncoder11
Model11
ResNetBackbone11
ResNetEmbeddings11
ResNetEncoder11
SinePositionEmbedding11

Config changes

9 fields changed, 71 unchanged.

FieldOriginalFixture
backbone_config.depths[3, 4, 6, 3][1, 1, 1, 1]
backbone_config.hidden_sizes[256, 512, 1024, 2048][16, 32, 64, 128]
d_model25664
decoder_attention_heads84
decoder_ffn_dim1024128
decoder_layers62
encoder_attention_heads84
encoder_ffn_dim1024128
encoder_in_channels[512, 1024, 2048][32, 64, 128]
\ No newline at end of file diff --git a/shrink/sam2.1-hiera-tiny-mic.html b/shrink/sam2.1-hiera-tiny-mic.html index 5c479ef91d690f5283e25498cd1da8e6b6d3c620..a7261c1d71505d5c4b736da98b3e382e23a2e4e2 100644 --- a/shrink/sam2.1-hiera-tiny-mic.html +++ b/shrink/sam2.1-hiera-tiny-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/sam2.1-hiera-tiny

huggingface-mic/sam2.1-hiera-tiny-mic@e740a31c3d86 ← facebook/sam2.1-hiera-tiny@de431c404385 · Sam2Model · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original31.4Mparameters
Fixture1.9Mparameters
Layer stacks3compared
Findings4to review

⚠ To review

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

vision_encoder.backbone.blocks

Original · 12
Fixture · 5
KindOriginalFixture
MultiScaleBlock125✓ kept

mask_decoder.transformer.layers

Original · 2
Fixture · 2
KindOriginalFixture
TwoWayAttentionBlock22✓ kept

mask_decoder.output_hypernetworks_mlps

Original · 4
Fixture · 4
KindOriginalFixture
ReLU44✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
vision_encoder.backbone.blocks26.8M85.3%1.2M61.2%
mask_decoder.transformer.layers3.2M10.0%398.3K21.0%
mask_decoder.output_hypernetworks_mlps559.2K1.8%140.4K7.4%
vision_encoder.neck369.7K1.2%62.0K3.3%
mask_decoder.iou_prediction_head132.6K0.4%33.5K1.8%
mask_decoder.transformer132.2K0.4%33.3K1.8%
mask_decoder.pred_obj_score_head131.8K0.4%33.2K1.7%
14 others131.7K0.4%37.3K2.0%

Preserved properties

ConfigPropertyOriginalFixture
modelRoPE typeaxialNone⚠ changed
modelmodel_typesam2_videosam2⚠ changed
modelimage_size10241024✓ kept
vision_configmodel_typesam2_vision_modelsam2_vision_model✓ kept
vision_confighidden_actgelugelu✓ kept
vision_config.backbone_configmodel_typesam2_hiera_det_modelsam2_hiera_det_model✓ kept
vision_config.backbone_configtie_word_embeddingsTrueTrue✓ kept
vision_config.backbone_confighidden_actgelugelu✓ kept
vision_config.backbone_configimage_size[1024, 1024][1024, 1024]✓ kept
vision_config.backbone_confignum_channels33✓ kept
prompt_encoder_configmodel_typesam2_video_prompt_encodersam2_prompt_encoder⚠ changed
prompt_encoder_confighidden_actgelugelu✓ kept
prompt_encoder_configimage_size10241024✓ kept
prompt_encoder_configpatch_size1616✓ kept
mask_decoder_configmodel_typesam2_video_mask_decodersam2_mask_decoder⚠ changed
mask_decoder_confighidden_actgelugelu✓ kept

Module classes

Instances of each class in the built model (prefix Sam2 omitted).

ClassOriginalFixture
Linear10173
LayerNorm3319
ModuleList2417
FeedForward2013
GELUActivation136
MultiScaleAttention125
MultiScaleBlock125
Conv2d1010
ReLU88
Attention77
Embedding66
LayerNorm33
ConvTranspose2d22
PositionalEmbedding22
TwoWayAttentionBlock22
GELU11
HieraDetModel11
MaskDecoder11
MaskEmbedding11
Model11
PatchEmbeddings11
PromptEncoder11
SinePositionEmbedding11
TwoWayTransformer11
VisionModel11
VisionNeck11

Config changes

18 fields changed, 134 unchanged.

FieldOriginalFixture
mask_decoder_config.hidden_size256128
mask_decoder_config.iou_head_hidden_dim256128
mask_decoder_config.mlp_dim2048256
mask_decoder_config.model_typesam2_video_mask_decodersam2_mask_decoder
mask_decoder_config.num_attention_heads84
max_cond_frame_num-1—
memory_attention_rope_theta—10000
model_typesam2_videosam2
prompt_encoder_config.hidden_size256128
prompt_encoder_config.model_typesam2_video_prompt_encodersam2_prompt_encoder
rope_parameters{"rope_theta": 10000, "rope_type": "axial"}—
vision_config.backbone_channel_list[768, 384, 192, 96][256, 128, 64, 32]
vision_config.backbone_config.blocks_per_stage[1, 2, 7, 2][1, 1, 2, 1]
vision_config.backbone_config.embed_dim_per_stage[96, 192, 384, 768][32, 64, 128, 256]
vision_config.backbone_config.global_attention_blocks[5, 7, 9][3]
vision_config.backbone_config.hidden_size9632
vision_config.backbone_config.num_attention_heads14
vision_config.fpn_hidden_size256128
\ No newline at end of file diff --git a/shrink/seamless-m4t-v2-large-mic.html b/shrink/seamless-m4t-v2-large-mic.html index b8d8e0d93073d5d0251873cd8027d9320db38242..b5904b146e5c266fce5bd9564b92471c3e835931 100644 --- a/shrink/seamless-m4t-v2-large-mic.html +++ b/shrink/seamless-m4t-v2-large-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/seamless-m4t-v2-large

huggingface-mic/seamless-m4t-v2-large-mic@43e20937aec2 ← facebook/seamless-m4t-v2-large@5f8cc790b19f · SeamlessM4Tv2Model · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original2.3Bparameters
Fixture95.0Mparameters
Layer stacks7compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

text_encoder.layers

Original · 24
Fixture · 2
KindOriginalFixture
EncoderLayer242✓ kept

speech_encoder.encoder.layers

Original · 24
Fixture · 24
KindOriginalFixture
ConformerEncoderLayer2424✓ kept

speech_encoder.adapter.layers

Original · 1
Fixture · 1
KindOriginalFixture
ConformerAdapterLayer11✓ kept

text_decoder.layers

Original · 24
Fixture · 2
KindOriginalFixture
DecoderLayer242✓ kept

t2u_model.model.encoder.layers

Original · 6
Fixture · 6
KindOriginalFixture
EncoderLayer66✓ kept

t2u_model.model.decoder.layers

Original · 6
Fixture · 6
KindOriginalFixture
TextToUnitDecoderLayer66✓ kept

vocoder.hifi_gan.resblocks

Original · 15
Fixture · 15
KindOriginalFixture
HifiGanResidualBlock1515✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
text_decoder.layers604.5M26.2%100.5K0.1%
speech_encoder.encoder.layers580.3M25.1%26.1M27.5%
text_encoder.layers503.7M21.8%66.9K0.1%
shared262.2M11.4%16.4M17.2%
t2u_model.model.encoder.layers125.9M5.5%6.4M6.8%
t2u_model.model.decoder.layers113.3M4.9%446.2K0.5%
speech_encoder.adapter.layers46.2M2.0%676.8K0.7%
13 others73.0M3.2%44.8M47.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeseamless_m4t_v2seamless_m4t_v2✓ kept
modelvocab_size256102256102✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelmax_position_embeddings40964096✓ kept
modelsampling_rate1600016000✓ kept

Module classes

Instances of each class in the built model (prefix SeamlessM4Tv2 omitted).

ClassOriginalFixture
Linear649297
Dropout346214
LayerNorm302192
Conv1d182182
ReLU8822
Attention8418
SiLU7272
FeedForwardNetwork5410
ConformerFeedForward5050
ModuleList3838
Embedding3030
EncoderLayer308
GLU2525
ConformerSelfAttention2525
ConformerConvolutionModule2424
ConformerEncoderLayer2424
DecoderLayer242
HifiGanResidualBlock1515
TextToUnitDecoderLayer66
ConvTranspose1d55
SinusoidalPositionalEmbedding44
Encoder22
ScaledWordEmbedding22
VariancePredictor22
CodeHifiGan11
ConformerAdapter11
ConformerAdapterLayer11
ConformerEncoder11
ConformerFeatureProjection11
Decoder11
HifiGan11
Model11
SpeechEncoder11
TextToUnitDecoder11
TextToUnitForConditionalGeneration11
TextToUnitModel11

Config changes

7 fields changed, 77 unchanged.

FieldOriginalFixture
decoder_attention_heads164
decoder_ffn_dim8192128
decoder_layers242
encoder_attention_heads164
encoder_ffn_dim8192128
encoder_layers242
hidden_size102464
\ No newline at end of file diff --git a/shrink/segformer-b0-finetuned-ade-512-512-mic.html b/shrink/segformer-b0-finetuned-ade-512-512-mic.html index 8be5defbdabeff6f379b761283f2e2fd438287db..722570d8bc4242fed590101c7b7b5b29c930d8d7 100644 --- a/shrink/segformer-b0-finetuned-ade-512-512-mic.html +++ b/shrink/segformer-b0-finetuned-ade-512-512-mic.html @@ -31,6 +31,7 @@ tr.bad td{color:var(--warn)} .c0{background:var(--s0)}.c1{background:var(--s1)}.c2{background:var(--s2)}.c3{background:var(--s3)}.c4{background:var(--s4)}.c5{background:var(--s5)}.c6{background:var(--s6)}.c7{background:var(--s7)}.c8{background:var(--s8)}

← MIC coverage of Transformers model types

-

nvidia/segformer-b0-finetuned-ade-512-512

huggingface-mic/segformer-b0-finetuned-ade-512-512-mic@b49b77daf875 ← nvidia/segformer-b0-finetuned-ade-512-512@489d5cd81a0b · SegformerForSemanticSegmentation · overfit weights · built before the shrink rule (no recorded recipe)

-
Original3.8Mparameters
Fixture787.3Kparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

segformer.stages

Original · 4
Fixture · 4
KindOriginalFixture
SequenceReduction33✓ kept
no SequenceReduction11✓ kept

decode_head.linear_projections

Original · 4
Fixture · 4
KindOriginalFixture
Linear44✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
segformer.stages3.3M88.5%423.6K53.8%
decode_head.linear_fuse262.1K7.0%262.1K33.3%
decode_head.linear_projections132.1K3.5%62.5K7.9%
decode_head.classifier38.5K1.0%38.5K4.9%
decode_head.batch_norm5120.0%5120.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typesegformersegformer✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size224224✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Segformer omitted).

ClassOriginalFixture
Linear5228
LayerNorm3019
Conv2d2013
Dropout179
GELUActivation84
Attention84
DepthWiseConv84
Layer84
MixMLP84
DropPath73
ModuleList66
SequenceReduction63
MLP44
OverlapPatchEmbeddings44
Stage44
BatchNorm2d11
Identity11
ReLU11
DecodeHead11
ForSemanticSegmentation11
Model11

Config changes

3 fields changed, 27 unchanged.

FieldOriginalFixture
depths[2, 2, 2, 2][1, 1, 1, 1]
hidden_sizes[32, 64, 160, 256][16, 32, 64, 128]
num_attention_heads[1, 2, 5, 8][1, 2, 4, 8]
+

nvidia/segformer-b0-finetuned-ade-512-512

huggingface-mic/segformer-b0-finetuned-ade-512-512-mic@6677cf3e6109 ← nvidia/segformer-b0-finetuned-ade-512-512@489d5cd81a0b · SegformerForSemanticSegmentation · real weights

+

CI: no failure on main in 30 days · MIC failures dashboard

+
Original3.8Mparameters
Fixture3.8Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

segformer.stages

Original · 4
Fixture · 4
KindOriginalFixture
SequenceReduction33✓ kept
no SequenceReduction11✓ kept

decode_head.linear_projections

Original · 4
Fixture · 4
KindOriginalFixture
Linear44✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
segformer.stages3.3M88.5%3.3M88.5%
decode_head.linear_fuse262.1K7.0%262.1K7.0%
decode_head.linear_projections132.1K3.5%132.1K3.5%
decode_head.classifier38.5K1.0%38.5K1.0%
decode_head.batch_norm5120.0%5120.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typesegformersegformer✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size224224✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Segformer omitted).

ClassOriginalFixture
Linear5252
LayerNorm3030
Conv2d2020
Dropout1717
GELUActivation88
Attention88
DepthWiseConv88
Layer88
MixMLP88
DropPath77
ModuleList66
SequenceReduction66
MLP44
OverlapPatchEmbeddings44
Stage44
BatchNorm2d11
Identity11
ReLU11
DecodeHead11
ForSemanticSegmentation11
Model11

Config changes

0 fields changed, 30 unchanged.

FieldOriginalFixture
\ No newline at end of file diff --git a/shrink/segformer-b0-finetuned-ade-512-512-mic.json b/shrink/segformer-b0-finetuned-ade-512-512-mic.json index 3b1d587539b7acc4aa5ff835eab1b16e6e64feb0..27eb4694d415daf5a2e53c1c3a8fe73934ae30f7 100644 --- a/shrink/segformer-b0-finetuned-ade-512-512-mic.json +++ b/shrink/segformer-b0-finetuned-ade-512-512-mic.json @@ -1 +1 @@ -{"fixture": "huggingface-mic/segformer-b0-finetuned-ade-512-512-mic@b49b77daf8758e1d1c6d7f37b58b535586bcf0fb", "findings": []} \ No newline at end of file +{"fixture": "huggingface-mic/segformer-b0-finetuned-ade-512-512-mic@6677cf3e61099c794f921bcd38ac95eb8a7e9ecf", "findings": []} \ No newline at end of file diff --git a/shrink/siglip-base-patch16-224-mic.html b/shrink/siglip-base-patch16-224-mic.html index 64f8ff9ffc7b578066bee8c054c54b8930a42bcb..94d1c3cff523328ee23fa204bebeae7c3a704ab0 100644 --- a/shrink/siglip-base-patch16-224-mic.html +++ b/shrink/siglip-base-patch16-224-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/siglip-base-patch16-224

huggingface-mic/siglip-base-patch16-224-mic@13fcabe2240b ← google/siglip-base-patch16-224@7fd15f0689c7 · SiglipModel · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original203.2Mparameters
Fixture2.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

text_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

vision_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
vision_model.encoder.layers85.1M41.9%66.9K2.9%
text_model.encoder.layers85.1M41.9%66.9K2.9%
text_model.embeddings24.6M12.1%2.1M89.8%
vision_model.head7.1M3.5%33.4K1.5%
vision_model.embeddings741.1K0.4%61.8K2.7%
text_model.head590.6K0.3%4.2K0.2%
vision_model.post_layernorm1.5K0.0%1280.0%
3 others1.5K0.0%1300.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typesiglipsiglip✓ kept
text_configmodel_typesiglip_text_modelsiglip_text_model✓ kept
text_configvocab_size3200032000✓ kept
text_confighidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
text_configmax_position_embeddings6464✓ kept
vision_configmodel_typesiglip_vision_modelsiglip_vision_model✓ kept
vision_confighidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
vision_configimage_size224224✓ kept
vision_configpatch_size1616✓ kept
vision_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Siglip omitted).

ClassOriginalFixture
Linear14727
LayerNorm5111
GELUTanh255
MLP255
Attention244
EncoderLayer244
Embedding33
ModuleList22
Encoder22
Conv2d11
MultiheadAttention11
NonDynamicallyQuantizableLinear11
Model11
MultiheadAttentionPoolingHead11
TextEmbeddings11
TextModel11
VisionEmbeddings11
VisionModel11

Config changes

9 fields changed, 42 unchanged.

FieldOriginalFixture
text_config.hidden_size76864
text_config.intermediate_size3072128
text_config.num_attention_heads124
text_config.num_hidden_layers122
text_config.projection_size76864
vision_config.hidden_size76864
vision_config.intermediate_size3072128
vision_config.num_attention_heads124
vision_config.num_hidden_layers122
\ No newline at end of file diff --git a/shrink/smollm-1.7b-mic.html b/shrink/smollm-1.7b-mic.html index 3bce35775a417605be913846f70939aa63ba1ba0..7d852412ea5e5757d17a2ac99cfbf9b677cfb6bf 100644 --- a/shrink/smollm-1.7b-mic.html +++ b/shrink/smollm-1.7b-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

HuggingFaceTB/SmolLM-1.7B

huggingface-mic/smollm-1.7b-mic@fc0e707f0977 ← HuggingFaceTB/SmolLM-1.7B@d7449ff7241c · LlamaForCausalLM · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original1.7Bparameters
Fixture3.2Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 24
Fixture · 2
KindOriginalFixture
DecoderLayer242✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers1.6B94.1%82.2K2.5%
model.embed_tokens100.7M5.9%3.1M97.5%
model.norm2.0K0.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleMHA (32 heads, 32 KV)MHA (4 heads, 4 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typellamallama✓ kept
modelvocab_size4915249152✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings20482048✓ kept

Module classes

Instances of each class in the built model (prefix Llama omitted).

ClassOriginalFixture
Linear16915
RMSNorm495
Attention242
DecoderLayer242
MLP242
SiLUActivation242
Embedding11
ForCausalLM11
Model11
RotaryEmbedding11
ModuleList11

Config changes

7 fields changed, 23 unchanged.

FieldOriginalFixture
head_dim6416
hidden_size204864
intermediate_size8192128
num_attention_heads324
num_hidden_layers242
num_key_value_heads324
pad_token_idnull0
\ No newline at end of file diff --git a/shrink/smollm-mic.html b/shrink/smollm-mic.html index 25ae4c89c082f9f440f53b75654e27016c828615..7d4c2d6f66f1279fc3ac347a7121fb3ecf372e06 100644 --- a/shrink/smollm-mic.html +++ b/shrink/smollm-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

HuggingFaceTB/SmolLM-135M

huggingface-mic/smollm-mic@d0da7b30e6c7 ← HuggingFaceTB/SmolLM-135M@1d461723eec6 · LlamaForCausalLM · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original134.5Mparameters
Fixture3.2Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 30
Fixture · 2
KindOriginalFixture
DecoderLayer302✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers106.2M79.0%74.0K2.3%
model.embed_tokens28.3M21.0%3.1M97.7%
model.norm5760.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleGQA (9 heads, 3 KV)GQA (4 heads, 2 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typellamallama✓ kept
modelvocab_size4915249152✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings20482048✓ kept

Module classes

Instances of each class in the built model (prefix Llama omitted).

ClassOriginalFixture
Linear21115
RMSNorm615
Attention302
DecoderLayer302
MLP302
SiLUActivation302
Embedding11
ForCausalLM11
Model11
RotaryEmbedding11
ModuleList11

Config changes

6 fields changed, 24 unchanged.

FieldOriginalFixture
head_dim6416
hidden_size57664
intermediate_size1536128
num_attention_heads94
num_hidden_layers302
num_key_value_heads32
\ No newline at end of file diff --git a/shrink/smolvlm-mic.html b/shrink/smolvlm-mic.html index 2e5489df468843c29f203c8a2e886e430d80157b..37402642c890cc2ebd5c534bf715dd6912d9b343 100644 --- a/shrink/smolvlm-mic.html +++ b/shrink/smolvlm-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

HuggingFaceTB/SmolVLM2-2.2B-Instruct

huggingface-mic/smolvlm-mic@8cefc6b12559 ← HuggingFaceTB/SmolVLM2-2.2B-Instruct@482adb537c02 · SmolVLMForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 15 failed checks in 30 days · MIC failures dashboard

Original2.2Bparameters
Fixture6.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.vision_model.encoder.layers

Original · 27
Fixture · 2
KindOriginalFixture
EncoderLayer272✓ kept

model.text_model.layers

Original · 24
Fixture · 2
KindOriginalFixture
LlamaDecoderLayer242✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.text_model.layers1.6B71.7%82.2K1.2%
model.vision_model.encoder.layers411.5M18.3%66.9K1.0%
model.text_model100.9M4.5%3.2M47.9%
lm_head100.9M4.5%3.2M47.9%
model.connector21.2M0.9%36.9K0.6%
model.vision_model1.5M0.1%84.5K1.3%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typesmolvlmsmolvlm✓ kept
modelvocab_size4928049280✓ kept
modeltie_word_embeddingsFalseFalse✓ kept
vision_configmodel_typesmolvlm_visionsmolvlm_vision✓ kept
vision_configtie_word_embeddingsFalseFalse✓ kept
vision_confighidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
vision_configimage_size384384✓ kept
vision_configpatch_size1414✓ kept
vision_confignum_channels33✓ kept
text_configattention styleMHA (32 heads, 32 KV)MHA (4 heads, 4 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typellamallama✓ kept
text_configvocab_size4928049280✓ kept
text_configtie_word_embeddingsFalseFalse✓ kept
text_confighidden_actsilusilu✓ kept
text_configmax_position_embeddings81928192✓ kept
text_config.perceiver_configmodel_typevllama3vllama3✓ kept
text_config.perceiver_configtie_word_embeddingsTrueTrue✓ kept
text_config.perceiver_confighidden_actsilusilu✓ kept

Module classes

Instances of each class in the built model (prefix SmolVLM omitted).

ClassOriginalFixture
Linear33228
LayerNorm555
LlamaRMSNorm495
GELUTanh272
EncoderLayer272
VisionAttention272
VisionMLP272
LlamaAttention242
LlamaDecoderLayer242
LlamaMLP242
SiLUActivation242
Embedding22
ModuleList22
Conv2d11
LlamaModel11
LlamaRotaryEmbedding11
Connector11
Encoder11
ForConditionalGeneration11
Model11
SimpleMLP11
VisionEmbeddings11
VisionTransformer11

Config changes

10 fields changed, 127 unchanged.

FieldOriginalFixture
text_config.head_dim6416
text_config.hidden_size204864
text_config.intermediate_size8192128
text_config.num_attention_heads324
text_config.num_hidden_layers242
text_config.num_key_value_heads324
vision_config.hidden_size115264
vision_config.intermediate_size4304128
vision_config.num_attention_heads164
vision_config.num_hidden_layers272
\ No newline at end of file diff --git a/shrink/speech_t5-mic.html b/shrink/speech_t5-mic.html index f5d5905f4d62bf8f5e144c0866406d5173973ff7..60a38b7a50cb09b7da8a1591ebbaa29873405d6f 100644 --- a/shrink/speech_t5-mic.html +++ b/shrink/speech_t5-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/speecht5_asr

huggingface-mic/speech_t5-mic@889b4e918f22 ← microsoft/speecht5_asr@53615c104084 · SpeechT5ForSpeechToText · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 3 failed checks in 30 days · MIC failures dashboard

Original151.2Mparameters
Fixture4.4Mparameters
Layer stacks3compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

speecht5.encoder.prenet.feature_encoder.conv_layers

Original · 7
Fixture · 7
KindOriginalFixture
GroupNorm, GroupNormConvLayer11✓ kept
NoLayerNormConvLayer66✓ kept

speecht5.encoder.wrapped_encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

speecht5.decoder.wrapped_decoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
DecoderLayer62✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
speecht5.encoder.wrapped_encoder.layers85.1M56.3%66.9K1.5%
speecht5.decoder.wrapped_decoder.layers56.7M37.5%100.5K2.3%
speecht5.encoder5.1M3.4%72.1K1.6%
speecht5.encoder.prenet.feature_encoder.conv_layers4.2M2.8%4.2M94.5%
speecht5.decoder62.2K0.0%5.2K0.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typespeecht5speecht5✓ kept
modelvocab_size8181✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelnum_mel_bins8080✓ kept

Module classes

Instances of each class in the built model (prefix SpeechT5 omitted).

ClassOriginalFixture
Linear13434
Dropout5715
LayerNorm4412
GELUActivation2612
Attention246
FeedForward184
EncoderLayer122
Conv1d77
DecoderLayer62
NoLayerNormConvLayer66
ModuleList33
Embedding22
SinusoidalPositionalEmbedding22
GroupNorm11
ModuleDict11
ParametrizationList11
ParametrizedConv1d11
Decoder11
DecoderWithTextPrenet11
Encoder11
EncoderWithSpeechPrenet11
FeatureEncoder11
FeatureProjection11
ForSpeechToText11
GroupNormConvLayer11
Model11
PositionalConvEmbedding11
RelativePositionalEncoding11
SamePadLayer11
SpeechEncoderPrenet11
TextDecoderPostnet11
TextDecoderPrenet11
_WeightNorm11

Config changes

7 fields changed, 60 unchanged.

FieldOriginalFixture
decoder_attention_heads124
decoder_ffn_dim3072128
decoder_layers62
encoder_attention_heads124
encoder_ffn_dim3072128
encoder_layers122
hidden_size76864
\ No newline at end of file diff --git a/shrink/stablelm-2-1_6b-mic.html b/shrink/stablelm-2-1_6b-mic.html index 2d9b60fdc147872faeefb38b7743c669643bf4c7..7ec0ebb432c8c409cfbd0dc0eabab35fada0253b 100644 --- a/shrink/stablelm-2-1_6b-mic.html +++ b/shrink/stablelm-2-1_6b-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

stabilityai/stablelm-2-1_6b

huggingface-mic/stablelm-2-1_6b-mic@eec0b53366b6 ← stabilityai/stablelm-2-1_6b@f499ead74c53 · StableLmForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original1.6Bparameters
Fixture12.9Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 24
Fixture · 2
KindOriginalFixture
DecoderLayer242✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers1.2B75.0%82.8K0.6%
model.embed_tokens205.5M12.5%6.4M49.7%
lm_head205.5M12.5%6.4M49.7%
model.norm4.1K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleMHA (32 heads, 32 KV)MHA (4 heads, 4 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typestablelmstablelm✓ kept
modelvocab_size100352100352✓ kept
modeltie_word_embeddingsFalseFalse✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings40964096✓ kept

Module classes

Instances of each class in the built model (prefix StableLm omitted).

ClassOriginalFixture
Linear16915
LayerNorm495
Dropout242
SiLUActivation242
Attention242
DecoderLayer242
MLP242
Embedding11
ModuleList11
ForCausalLM11
Model11
RotaryEmbedding11

Config changes

5 fields changed, 25 unchanged.

FieldOriginalFixture
hidden_size204864
intermediate_size5632128
num_attention_heads324
num_hidden_layers242
num_key_value_heads324
\ No newline at end of file diff --git a/shrink/starcoder2-3b-mic.html b/shrink/starcoder2-3b-mic.html index 0bb6947f812d0564a7d82bfe0f893496394e1542..0e44f6e8c644e9c4d9f12274c28db009a671adf2 100644 --- a/shrink/starcoder2-3b-mic.html +++ b/shrink/starcoder2-3b-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

bigcode/starcoder2-3b

huggingface-mic/starcoder2-3b-mic@f158c5574331 ← bigcode/starcoder2-3b@733247c55e3f · Starcoder2ForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original3.0Bparameters
Fixture3.2Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 30
Fixture · 2
KindOriginalFixture
DecoderLayer302✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.layers2.9B95.0%58.6K1.8%
model.embed_tokens151.0M5.0%3.1M98.2%
model.norm6.1K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleGQA (24 heads, 2 KV)GQA (4 heads, 2 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typestarcoder2starcoder2✓ kept
modelvocab_size4915249152✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
modelmax_position_embeddings1638416384✓ kept

Module classes

Instances of each class in the built model (prefix Starcoder2 omitted).

ClassOriginalFixture
Linear18113
LayerNorm615
GELUTanh302
Attention302
DecoderLayer302
MLP302
Embedding11
ModuleList11
ForCausalLM11
Model11
RotaryEmbedding11

Config changes

4 fields changed, 28 unchanged.

FieldOriginalFixture
hidden_size307264
intermediate_size12288128
num_attention_heads244
num_hidden_layers302
\ No newline at end of file diff --git a/shrink/swin-tiny-patch4-window7-224-mic.html b/shrink/swin-tiny-patch4-window7-224-mic.html index 8959797b2d2a7ad01bebec2c5f60dd7a1f04faa2..125f6d9736fd3ee26dddafb9cb502a7d98fcb0c7 100644 --- a/shrink/swin-tiny-patch4-window7-224-mic.html +++ b/shrink/swin-tiny-patch4-window7-224-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/swin-tiny-patch4-window7-224

huggingface-mic/swin-tiny-patch4-window7-224-mic@89ae45d56ca2 ← microsoft/swin-tiny-patch4-window7-224@d00d478bfaf1 · SwinForImageClassification · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original28.3Mparameters
Fixture5.4Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

swin.encoder.layers

Original · 4
Fixture · 4
KindOriginalFixture
PatchMerging33✓ kept
no PatchMerging11✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
swin.encoder.layers27.5M97.3%4.9M90.4%
classifier769.0K2.7%513.0K9.5%
swin.embeddings4.9K0.0%3.3K0.1%
swin.layernorm1.5K0.0%1.0K0.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeswinswin✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size224224✓ kept
modelpatch_size44✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Swin omitted).

ClassOriginalFixture
Linear7628
LayerNorm2913
Dropout135
GELUActivation124
Attention124
Layer124
MLP124
RelativePositionBias124
DropPath113
ModuleList55
Stage44
PatchMerging33
AdaptiveAvgPool1d11
Conv2d11
Identity11
Embeddings11
Encoder11
ForImageClassification11
Model11
PatchEmbeddings11

Config changes

5 fields changed, 27 unchanged.

FieldOriginalFixture
depths[2, 2, 6, 2][1, 1, 1, 1]
embed_dim9664
hidden_size76864
num_heads[3, 6, 12, 24][2, 4, 8, 16]
num_layers42
\ No newline at end of file diff --git a/shrink/swinv2-tiny-patch4-window16-256-mic.html b/shrink/swinv2-tiny-patch4-window16-256-mic.html index 776deb410947023670929d32e8e898e62320ec09..4507a1708170db6533eb97dc326e5b7db74446a1 100644 --- a/shrink/swinv2-tiny-patch4-window16-256-mic.html +++ b/shrink/swinv2-tiny-patch4-window16-256-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/swinv2-tiny-patch4-window16-256

huggingface-mic/swinv2-tiny-patch4-window16-256-mic@8c997579b76f ← microsoft/swinv2-tiny-patch4-window16-256@f4d3075206f2 · Swinv2ForImageClassification · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original28.3Mparameters
Fixture5.4Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

swinv2.encoder.layers

Original · 4
Fixture · 4
KindOriginalFixture
PatchMerging33✓ kept
no PatchMerging11✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
swinv2.encoder.layers27.6M97.3%4.9M90.5%
classifier769.0K2.7%513.0K9.5%
swinv2.embeddings4.9K0.0%3.3K0.1%
swinv2.layernorm1.5K0.0%1.0K0.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeswinv2swinv2✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size256256✓ kept
modelpatch_size44✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Swinv2 omitted).

ClassOriginalFixture
Linear10036
Dropout3713
LayerNorm2913
GELUActivation124
ReLU124
Sequential124
Attention124
Intermediate124
Layer124
Output124
SelfAttention124
SelfOutput124
DropPath113
ModuleList55
Stage44
PatchMerging33
AdaptiveAvgPool1d11
Conv2d11
Identity11
Embeddings11
Encoder11
ForImageClassification11
Model11
PatchEmbeddings11

Config changes

5 fields changed, 28 unchanged.

FieldOriginalFixture
depths[2, 2, 6, 2][1, 1, 1, 1]
embed_dim9664
hidden_size76864
num_heads[3, 6, 12, 24][2, 4, 8, 16]
num_layers42
\ No newline at end of file diff --git a/shrink/t5-mic.html b/shrink/t5-mic.html index 2988b1dddb3b2aaaad07c731fea45f94d2dd9868..420a06798b7ea9886bd5e0a565b313147aad3b20 100644 --- a/shrink/t5-mic.html +++ b/shrink/t5-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google-t5/t5-small

huggingface-mic/t5-mic@17a8c0b6ac87 ← google-t5/t5-small@df1b051c4962 · T5ForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 3 failed checks in 30 days · MIC failures dashboard

Original60.5Mparameters
Fixture2.2Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

encoder.block

Original · 6
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding51✓ kept

decoder.block

Original · 6
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding51✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
decoder.block25.2M41.6%98.8K4.4%
encoder.block18.9M31.2%65.9K3.0%
shared16.4M27.2%2.1M92.6%
decoder.final_layer_norm5120.0%640.0%
encoder.final_layer_norm5120.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typet5t5✓ kept
modelvocab_size3212832128✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix T5 omitted).

ClassOriginalFixture
Linear9733
Dropout4416
LayerNorm3212
Attention186
ModuleList146
ReLU124
Block124
DenseActDense124
LayerFF124
LayerSelfAttention124
LayerCrossAttention62
Embedding55
Stack22
ForConditionalGeneration11

Config changes

6 fields changed, 29 unchanged.

FieldOriginalFixture
d_ff2048128
d_kv6416
d_model51264
num_decoder_layers62
num_heads84
num_layers62
\ No newline at end of file diff --git a/shrink/t5-small-mic.html b/shrink/t5-small-mic.html index e79d21d5f361f9ac4696adbe5aec406d61466bd1..f2ddf53e8373d67cf14441e663312bc879927bae 100644 --- a/shrink/t5-small-mic.html +++ b/shrink/t5-small-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google-t5/t5-small

huggingface-mic/t5-small-mic@4f5dcdc60406 ← google-t5/t5-small@df1b051c4962 · T5ForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original60.5Mparameters
Fixture2.2Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

encoder.block

Original · 6
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding51✓ kept

decoder.block

Original · 6
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding51✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
decoder.block25.2M41.6%98.8K4.4%
encoder.block18.9M31.2%65.9K3.0%
shared16.4M27.2%2.1M92.6%
decoder.final_layer_norm5120.0%640.0%
encoder.final_layer_norm5120.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typet5t5✓ kept
modelvocab_size3212832128✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix T5 omitted).

ClassOriginalFixture
Linear9733
Dropout4416
LayerNorm3212
Attention186
ModuleList146
ReLU124
Block124
DenseActDense124
LayerFF124
LayerSelfAttention124
LayerCrossAttention62
Embedding55
Stack22
ForConditionalGeneration11

Config changes

6 fields changed, 29 unchanged.

FieldOriginalFixture
d_ff2048128
d_kv6416
d_model51264
num_decoder_layers62
num_heads84
num_layers62
\ No newline at end of file diff --git a/shrink/table_transformer-mic.html b/shrink/table_transformer-mic.html index 21a4b72628eeafd1dff13adb5b0c8ee3e72e0ed0..2568ead2edf77c7cdd3ba809952aee2e9a70be1c 100644 --- a/shrink/table_transformer-mic.html +++ b/shrink/table_transformer-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/table-transformer-detection

huggingface-mic/table_transformer-mic@5b726b4a1f1b ← microsoft/table-transformer-detection@2357cbe2b5a5 · TableTransformerForObjectDetection · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 3 failed checks in 30 days · MIC failures dashboard

Original28.8Mparameters
Fixture11.4Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
EncoderLayer62✓ kept

model.decoder.layers

Original · 6
Fixture · 2
KindOriginalFixture
DecoderLayer62✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.backbone11.2M38.8%11.2M98.2%
model.decoder.layers9.5M32.9%100.5K0.9%
model.encoder.layers7.9M27.4%66.9K0.6%
bbox_predictor.layers132.6K0.5%8.6K0.1%
model.input_projection131.3K0.5%32.8K0.3%
model.query_position_embeddings3.8K0.0%9600.0%
class_labels_classifier7710.0%1950.0%
2 others1.0K0.0%2560.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typetable-transformertable-transformer✓ kept
modelmax_position_embeddings10241024✓ kept
modelnum_channels33✓ kept
backbone_configmodel_typetimm_backbonetimm_backbone✓ kept
backbone_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix TableTransformer omitted).

ClassOriginalFixture
Linear10036
LayerNorm3212
ReLU2921
Conv2d2121
FrozenBatchNorm2d2020
Attention186
Identity1616
BasicBlock88
Sequential77
DecoderLayer62
EncoderLayer62
ModuleList33
Embedding11
FeatureListNet11
MaxPool2d11
ConvEncoder11
ConvModel11
Decoder11
Encoder11
ForObjectDetection11
MLPPredictionHead11
Model11
SinePositionEmbedding11

Config changes

10 fields changed, 48 unchanged.

FieldOriginalFixture
backbone_config.out_featuresnull["layer4"]
backbone_config.out_indices[1, 2, 3, 4][-1]
backbone_config.stage_names—["act1", "layer1", "layer2", "layer3", "layer4"]
d_model25664
decoder_attention_heads84
decoder_ffn_dim2048128
decoder_layers62
encoder_attention_heads84
encoder_ffn_dim2048128
encoder_layers62
\ No newline at end of file diff --git a/shrink/tibetan-ocr-mic.html b/shrink/tibetan-ocr-mic.html index e97f774cbf860a96dede1edbe93d479d4ae1a9f5..b412e418c669d7791c20a61dbbd5b461cb00a770 100644 --- a/shrink/tibetan-ocr-mic.html +++ b/shrink/tibetan-ocr-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

BDRC/tibetan-ocr

huggingface-mic/tibetan-ocr-mic@78b84a91abf7 ← BDRC/tibetan-ocr@50506eb6d8ed · PaddleOCRVLForConditionalGeneration · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original810.7Mparameters
Fixture535.7Kparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.visual.vision_model.encoder.layers

Original · 27
Fixture · 2
KindOriginalFixture
PaddleOCRVisionEncoderLayer272✓ kept

model.language_model.layers

Original · 26
Fixture · 2
KindOriginalFixture
PaddleOCRDecoderLayer262✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.visual.vision_model.encoder.layers411.5M50.8%66.9K12.5%
model.language_model.layers368.1M45.4%74.0K13.8%
model.projector26.0M3.2%82.4K15.4%
model.language_model3.6M0.4%227.9K42.5%
model.visual1.5M0.2%84.5K15.8%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typepaddleocr_vlpaddleocr_vl✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
text_configattention styleGQA (16 heads, 2 KV)GQA (4 heads, 2 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typepaddleocr_vl_textpaddleocr_vl_text✓ kept
text_configvocab_size35603560✓ kept
text_configtie_word_embeddingsTrueTrue✓ kept
text_confighidden_actsilusilu✓ kept
text_configmax_position_embeddings131072131072✓ kept
vision_configRoPE typeaxialaxial✓ kept
vision_configmodel_typepaddleocr_vl_visionpaddleocr_vl_vision✓ kept
vision_confighidden_actgelu_pytorch_tanhgelu_pytorch_tanh✓ kept
vision_configimage_size384384✓ kept
vision_configpatch_size1414✓ kept
vision_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix PaddleOCRVL omitted).

ClassOriginalFixture
Linear34729
LayerNorm566
PaddleOCRRMSNorm535
GELUTanh272
PaddleOCRVisionAttention272
PaddleOCRVisionEncoderLayer272
PaddleOCRVisionMLP272
PaddleOCRAttention262
PaddleOCRDecoderLayer262
PaddleOCRMLP262
SiLUActivation262
Embedding22
ModuleList22
Conv2d11
GELUActivation11
PaddleOCRProjector11
PaddleOCRRotaryEmbedding11
PaddleOCRTextModel11
ForConditionalGeneration11
Model11
PaddleOCRVisionEmbeddings11
PaddleOCRVisionEncoder11
PaddleOCRVisionModel11
PaddleOCRVisionRotaryEmbedding11
PaddleOCRVisionTransformer11

Config changes

10 fields changed, 66 unchanged.

FieldOriginalFixture
text_config.head_dim12816
text_config.hidden_size102464
text_config.intermediate_size3072128
text_config.num_attention_heads164
text_config.num_hidden_layers262
text_config.rope_parameters{"mrope_section": [16, 24, 24], "rope_theta": 500000, "rope_type": "default",…{"mrope_section": [2, 3, 3], "rope_theta": 500000, "rope_type": "default", "t…
vision_config.hidden_size115264
vision_config.intermediate_size4304128
vision_config.num_attention_heads164
vision_config.num_hidden_layers272
\ No newline at end of file diff --git a/shrink/timesfm-2.0-500m-pytorch-mic.html b/shrink/timesfm-2.0-500m-pytorch-mic.html index 4cafe39de66c845a37fc5b129a5de531ed6e6a9c..0fc87cdda7fdc003a43054b1fb7f98a545d6f7b0 100644 --- a/shrink/timesfm-2.0-500m-pytorch-mic.html +++ b/shrink/timesfm-2.0-500m-pytorch-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/timesfm-2.0-500m-pytorch

huggingface-mic/timesfm-2.0-500m-pytorch-mic@d9e2f9e3ea41 ← google/timesfm-2.0-500m-pytorch@dc2443792ce5 · TimesFmModelForPrediction · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original498.8Mparameters
Fixture344.4Kparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

decoder.layers

Original · 50
Fixture · 2
KindOriginalFixture
DecoderLayer502✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
decoder.layers492.1M98.7%66.8K19.4%
decoder.input_ff_layer1.8M0.4%20.7K6.0%
horizon_ff_layer.output_layer1.6M0.3%165.1K47.9%
horizon_ff_layer.residual_layer1.6M0.3%83.2K24.2%
horizon_ff_layer.input_layer1.6M0.3%8.3K2.4%
decoder.freq_emb3.8K0.0%1920.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typetimesfmtimesfm✓ kept

Module classes

Instances of each class in the built model (prefix TimesFm omitted).

ClassOriginalFixture
Linear30618
LayerNorm502
Attention502
DecoderLayer502
MLP502
RMSNorm502
SiLU22
ResidualBlock22
Embedding11
ModuleList11
Model11
ModelForPrediction11

Config changes

5 fields changed, 22 unchanged.

FieldOriginalFixture
head_dim8016
hidden_size128064
intermediate_size1280128
num_attention_heads164
num_hidden_layers502
\ No newline at end of file diff --git a/shrink/tiny-clip-text-2-mic.html b/shrink/tiny-clip-text-2-mic.html index 82e8a1743ea8bc74b0ab8f9d3e4d8b63dda64b88..78f7b2a3aa884bd850add06ade58c7663144766b 100644 --- a/shrink/tiny-clip-text-2-mic.html +++ b/shrink/tiny-clip-text-2-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

peft-internal-testing/tiny-clip-text-2

huggingface-mic/tiny-clip-text-2-mic@c9fce363d542 ← peft-internal-testing/tiny-clip-text-2@4fbb8d27f0fd · CLIPTextModelWithProjection · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original69.5Kparameters
Fixture138.0Kparameters
Layer stacks1compared
Findings2to review

⚠ To review

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

text_model.encoder.layers

Original · 5
Fixture · 2
KindOriginalFixture
CLIPEncoderLayer52✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
text_model.embeddings34.5K49.6%68.9K49.9%
text_model.encoder.layers33.9K48.8%66.9K48.5%
text_projection1.0K1.5%2.0K1.5%
text_model.final_layer_norm640.1%1280.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeclip_text_modelclip_text_model✓ kept
modelvocab_size10001000✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings7777✓ kept

Module classes

Instances of each class in the built model (prefix CLIPText omitted).

ClassOriginalFixture
Linear3113
LayerNorm115
CLIPAttention52
CLIPEncoderLayer52
CLIPMLP52
GELUActivation52
Embedding22
CLIPEncoder11
Embeddings11
Model11
ModelWithProjection11
ModuleList11

Config changes

3 fields changed, 21 unchanged.

FieldOriginalFixture
hidden_size3264
intermediate_size37128
num_hidden_layers52
\ No newline at end of file diff --git a/shrink/tiny-random-aya-base-mic.html b/shrink/tiny-random-aya-base-mic.html index 60fd43cdc827c0720cd2c4688fce92889aad586d..ee132e107c8064fd4a9ac897aa5b58d0ba2d7289 100644 --- a/shrink/tiny-random-aya-base-mic.html +++ b/shrink/tiny-random-aya-base-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

optimum-intel-internal-testing/tiny-random-aya-base

huggingface-mic/tiny-random-aya-base-mic@42e97ac577e1 ← optimum-intel-internal-testing/tiny-random-aya-base@e85478d37f91 · Cohere2ForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original1.1Mparameters
Fixture16.9Mparameters
Layer stacks1compared
Findings2to review

⚠ To review

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.layers

Original · 2
Fixture · 2
KindOriginalFixture
DecoderLayer22✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.embed_tokens1.0M98.4%16.8M99.5%
model.layers16.5K1.6%82.0K0.5%
model.norm40.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelattention styleMHA (4 heads, 4 KV)MHA (4 heads, 4 KV)✓ kept
modelRoPE typedefaultdefault✓ kept
modelmodel_typecohere2cohere2✓ kept
modelvocab_size262144262144✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actsilusilu✓ kept
modelmax_position_embeddings500000500000✓ kept

Module classes

Instances of each class in the built model (prefix Cohere2 omitted).

ClassOriginalFixture
Linear1515
LayerNorm33
Attention22
DecoderLayer22
MLP22
SiLUActivation22
ForCausalLM11
Model11
RotaryEmbedding11
Embedding11
ModuleList11

Config changes

3 fields changed, 40 unchanged.

FieldOriginalFixture
head_dim12816
hidden_size464
intermediate_size5128
\ No newline at end of file diff --git a/shrink/tiny-random-muse-glimmer-mic.html b/shrink/tiny-random-muse-glimmer-mic.html index 65fe1f53419e2f3a7d4ea42dfcffaf3e95b6a225..569e033ab19f0efb9ac35b65a9e9f5cd6d5d2807 100644 --- a/shrink/tiny-random-muse-glimmer-mic.html +++ b/shrink/tiny-random-muse-glimmer-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

optimum-intel-internal-testing/tiny-random-muse-glimmer

huggingface-mic/tiny-random-muse-glimmer-mic@5d21eb327b89 ← optimum-intel-internal-testing/tiny-random-muse-glimmer@cb70af191bbb · MuseGlimmerForConditionalGeneration · random weights

+

CI: failing now on main (1 failed checks in 30 days) · MIC failures dashboard

Original13.5Mparameters
Fixture13.3Mparameters
Layer stacks2compared
Findings2to review

⚠ To review

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.vision_tower.layers

Original · 4
Fixture · 2
KindOriginalFixture
VisionEncoderLayer42✓ kept

model.language_model.layers

Original · 4
Fixture · 2
KindOriginalFixture
TextDecoderLayer42✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.language_model12.9M95.6%12.9M97.4%
model.language_model.layers263.2K1.9%82.4K0.6%
model.vision_tower141.1K1.0%141.1K1.1%
model.vision_tower.layers133.9K1.0%66.9K0.5%
model.vision_adapter49.2K0.4%49.2K0.4%
model.vision_projection8.2K0.1%8.2K0.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typemuse_glimmermuse_glimmer✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
text_configattention styleGQA (4 heads, 2 KV)GQA (4 heads, 2 KV)✓ kept
text_configRoPE typedefaultdefault✓ kept
text_configmodel_typemuse_glimmer_textmuse_glimmer_text✓ kept
text_configvocab_size202048202048✓ kept
text_configtie_word_embeddingsTrueTrue✓ kept
text_configmax_position_embeddings131072131072✓ kept
vision_configRoPE typeaxialaxial✓ kept
vision_configmodel_typemuse_glimmer_visionmuse_glimmer_vision✓ kept
vision_confighidden_actgelugelu✓ kept
vision_configmax_position_embeddings10241024✓ kept
vision_configpatch_size1414✓ kept

Module classes

Instances of each class in the built model (prefix MuseGlimmer omitted).

ClassOriginalFixture
Linear6133
TextCenteredRMSNorm168
LayerNorm106
RMSNorm75
GELUActivation53
TextAttention42
TextDecoderLayer42
TextMLP42
VisionAttention42
VisionEncoderLayer42
VisionMLP42
SiLUActivation42
ModuleList22
Embedding11
ForConditionalGeneration11
Model11
TextModel11
TextNormedEmbedding11
TextRotaryEmbedding11
VisionAdapter11
VisionModel11
VisionPatchEmbedder11
VisionRotaryEmbedding11

Config changes

6 fields changed, 67 unchanged.

FieldOriginalFixture
text_config.intermediate_size256128
text_config.layer_rope_theta[500000.0, 500000.0, 500000.0, 0][500000.0, 0]
text_config.layer_types["sliding_attention", "sliding_attention", "sliding_attention", "full_attenti…["sliding_attention", "full_attention"]
text_config.num_hidden_layers42
vision_config.layer_types["window_attention", "window_attention", "window_attention", "full_attention"]["window_attention", "full_attention"]
vision_config.num_hidden_layers42
\ No newline at end of file diff --git a/shrink/tiny_starcoder_py-mic.html b/shrink/tiny_starcoder_py-mic.html index 3909c8c85581312f5341f4e8243626091ec77342..8cf3136f8b8018edb094314bb2da533fe1c8b32c 100644 --- a/shrink/tiny_starcoder_py-mic.html +++ b/shrink/tiny_starcoder_py-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

bigcode/tiny_starcoder_py

huggingface-mic/tiny_starcoder_py-mic@0ad90af0e2e2 ← bigcode/tiny_starcoder_py@8547527bef0b · GPTBigCodeForCausalLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original164.1Mparameters
Fixture3.7Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

transformer.h

Original · 20
Fixture · 2
KindOriginalFixture
Block202✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
transformer.h120.1M73.2%54.5K1.5%
transformer.wte37.7M23.0%3.1M84.5%
transformer.wpe6.3M3.8%524.3K14.1%
transformer.ln_f1.5K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typegpt_bigcodegpt_bigcode✓ kept
modelvocab_size4915249152✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix GPTBigCode omitted).

ClassOriginalFixture
Linear819
Dropout415
LayerNorm415
GELUTanh202
Attention202
Block202
MLP202
Embedding22
ForCausalLM11
Model11
ModuleList11

Config changes

5 fields changed, 38 unchanged.

FieldOriginalFixture
n_embd76864
n_head124
n_inner3072128
n_layer202
num_key_value_heads12
\ No newline at end of file diff --git a/shrink/tipsv2-b14-mic.html b/shrink/tipsv2-b14-mic.html index 3d4bfb88b1c7c09c51e2c8ac9022bc0d349ed959..10013b4e2bab28b5410772bd92e9f04946fc91fe 100644 --- a/shrink/tipsv2-b14-mic.html +++ b/shrink/tipsv2-b14-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/tipsv2-b14

huggingface-mic/tipsv2-b14-mic@1ba3c8e34558 ← google/tipsv2-b14@ed1e4dc6b74b · Tipsv2Model · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original195.9Mparameters
Fixture2.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

text_model.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
TextEncoderLayer122✓ kept

vision_model.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
VisionLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
vision_model.encoder.layer85.1M43.4%100.2K4.3%
text_model.encoder.layers85.1M43.4%66.9K2.9%
text_model.embeddings24.6M12.5%2.0M88.3%
vision_model.embeddings1.2M0.6%103.5K4.5%
text_model.final_layer_norm1.5K0.0%1280.0%
vision_model.layernorm1.5K0.0%1280.0%
temperature10.0%10.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typetipsv2tipsv2✓ kept
text_configmodel_typetipsv2_text_modeltipsv2_text_model✓ kept
text_configvocab_size3200032000✓ kept
text_confighidden_actrelurelu✓ kept
text_configmax_position_embeddings6464✓ kept
vision_configmodel_typetipsv2_vision_modeltipsv2_vision_model✓ kept
vision_confighidden_actgelugelu✓ kept
vision_configimage_size448448✓ kept
vision_configpatch_size1414✓ kept
vision_confignum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Tipsv2 omitted).

ClassOriginalFixture
Linear14424
LayerNorm5010
VisionLayerScale244
GELUActivation122
Identity122
ReLU122
MLP122
TextAttention122
TextEncoderLayer122
VisionAttention122
VisionLayer122
VisionMLP122
ModuleList22
Conv2d11
Dropout11
Embedding11
Model11
SinusoidalPositionalEmbedding11
TextEmbeddings11
TextEncoder11
TextModel11
VisionEmbeddings11
VisionEncoder11
VisionModel11
VisionPatchEmbeddings11

Config changes

9 fields changed, 59 unchanged.

FieldOriginalFixture
text_config.hidden_size76864
text_config.intermediate_size3072128
text_config.num_attention_heads124
text_config.num_hidden_layers122
vision_config.hidden_size76864
vision_config.num_attention_heads124
vision_config.num_hidden_layers122
vision_config.out_features["stage12"]["stage2"]
vision_config.out_indices[12][2]
\ No newline at end of file diff --git a/shrink/umt5-small-mic.html b/shrink/umt5-small-mic.html index 481cac289807c5342921a97d2d3591e8a50a7faa..94df69e192912230f6889c088f5af330e1fb3054 100644 --- a/shrink/umt5-small-mic.html +++ b/shrink/umt5-small-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/umt5-small

huggingface-mic/umt5-small-mic@3ea88ecd4cac ← google/umt5-small@8c63c2b77efb · UMT5ForConditionalGeneration · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original175.3Mparameters
Fixture16.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

encoder.block

Original · 8
Fixture · 2
KindOriginalFixture
Block82✓ kept

decoder.block

Original · 8
Fixture · 2
KindOriginalFixture
Block82✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
shared131.3M74.9%16.4M98.8%
decoder.block25.2M14.4%115.3K0.7%
encoder.block18.9M10.8%82.4K0.5%
decoder.final_layer_norm5120.0%640.0%
encoder.final_layer_norm5120.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeumt5umt5✓ kept
modelvocab_size256384256384✓ kept
modeltie_word_embeddingsTrueTrue✓ kept

Module classes

Instances of each class in the built model (prefix UMT5 omitted).

ClassOriginalFixture
Linear14537
Dropout5816
LayerNorm4212
Attention246
Embedding197
ModuleList186
NewGELUActivation164
Block164
DenseGatedActDense164
LayerFF164
LayerSelfAttention164
LayerCrossAttention82
Stack22
ForConditionalGeneration11

Config changes

6 fields changed, 28 unchanged.

FieldOriginalFixture
d_ff1024128
d_kv6416
d_model51264
num_decoder_layers82
num_heads64
num_layers82
\ No newline at end of file diff --git a/shrink/vit-mic.html b/shrink/vit-mic.html index 41f99fea4b33279f06d2dc231907832b5f7f2776..c98b3163b4c47b7bced30fa70ced8a029a19edd7 100644 --- a/shrink/vit-mic.html +++ b/shrink/vit-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

google/vit-base-patch16-224

huggingface-mic/vit-mic@d69cbc45edf9 ← google/vit-base-patch16-224@3f49326eb077 · ViTForImageClassification · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 3 failed checks in 30 days · MIC failures dashboard

Original86.6Mparameters
Fixture194.0Kparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

vit.layers

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
vit.layers85.1M98.3%66.9K34.5%
classifier769.0K0.9%65.0K33.5%
vit.embeddings742.7K0.9%61.9K31.9%
vit.layernorm1.5K0.0%1280.1%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typevitvit✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size224224✓ kept
modelpatch_size1616✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix ViT omitted).

ClassOriginalFixture
Linear7313
LayerNorm255
Dropout133
GELUActivation122
Attention122
Layer122
MLP122
Conv2d11
ModuleList11
Embeddings11
ForImageClassification11
Model11
PatchEmbeddings11

Config changes

5 fields changed, 20 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
problem_typenullsingle_label_classification
\ No newline at end of file diff --git a/shrink/vitmatte-small-composition-1k-mic.html b/shrink/vitmatte-small-composition-1k-mic.html index d826a3c2e0cf44eea53014a07f418be3aeefdc8b..be2307a31ca26738566d5ffff5598b885af33ca1 100644 --- a/shrink/vitmatte-small-composition-1k-mic.html +++ b/shrink/vitmatte-small-composition-1k-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

hustvl/vitmatte-small-composition-1k

huggingface-mic/vitmatte-small-composition-1k-mic@61792a274450 ← hustvl/vitmatte-small-composition-1k@6a58ad764640 · VitMatteForImageMatting · overfit weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original25.8Mparameters
Fixture1.6Mparameters
Layer stacks3compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

backbone.encoder.layer

Original · 12
Fixture · 3
KindOriginalFixture
no Conv2d82✓ kept
Conv2d, VitDetLayerNorm, VitDetResBottleneckBlock41✓ kept

decoder.convstream.convs

Original · 3
Fixture · 3
KindOriginalFixture
ReLU33✓ kept

decoder.fusion_blocks

Original · 4
Fixture · 4
KindOriginalFixture
FusionBlock44✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
backbone.encoder.layer23.3M90.2%167.3K10.6%
decoder.fusion_blocks1.9M7.2%1.1M70.8%
backbone.embeddings469.2K1.8%78.2K5.0%
decoder.convstream.convs209.8K0.8%209.8K13.3%
decoder.matting_head4.7K0.0%4.7K0.3%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typevitmattevitmatte✓ kept
backbone_configmodel_typevitdetvitdet✓ kept
backbone_confighidden_actgelugelu✓ kept
backbone_configimage_size512512✓ kept
backbone_configpatch_size1616✓ kept
backbone_confignum_channels44✓ kept

Module classes

Instances of each class in the built model (prefix VitMatte omitted).

ClassOriginalFixture
Linear4812
LayerNorm246
Conv2d2213
GELUActivation205
Dropout123
Identity123
VitDetAttention123
VitDetLayer123
VitDetLayerNorm123
VitDetMlp123
BatchNorm2d88
ReLU88
BasicConv3x377
VitDetResBottleneckBlock41
FusionBlock44
ModuleList33
Sequential11
VitDetBackbone11
VitDetEmbeddings11
VitDetEncoder11
ConvStream11
DetailCaptureModule11
ForImageMatting11
Head11

Config changes

6 fields changed, 39 unchanged.

FieldOriginalFixture
backbone_config.hidden_size38464
backbone_config.num_attention_heads64
backbone_config.num_hidden_layers123
backbone_config.out_features["stage12"]["stage3"]
backbone_config.out_indices[12][3]
hidden_size38464
\ No newline at end of file diff --git a/shrink/wav2vec2-base-960h-mic.html b/shrink/wav2vec2-base-960h-mic.html index e6aeacad43245264f53ae11d7d0104331617e0dd..15554d4927e7c20db27d8b9bcfdac4b4def009cd 100644 --- a/shrink/wav2vec2-base-960h-mic.html +++ b/shrink/wav2vec2-base-960h-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/wav2vec2-base-960h

huggingface-mic/wav2vec2-base-960h-mic@889c4c48f746 ← facebook/wav2vec2-base-960h@22aad52d435e · Wav2Vec2ForCTC · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original94.4Mparameters
Fixture4.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

wav2vec2.feature_extractor.conv_layers

Original · 7
Fixture · 7
KindOriginalFixture
GroupNorm, GroupNormConvLayer11✓ kept
NoLayerNormConvLayer66✓ kept

wav2vec2.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
wav2vec2.encoder.layers85.1M90.1%66.9K1.5%
wav2vec2.encoder4.7M5.0%33.1K0.8%
wav2vec2.feature_extractor.conv_layers4.2M4.4%4.2M96.9%
wav2vec2.feature_projection395.0K0.4%33.9K0.8%
lm_head24.6K0.0%2.1K0.0%
wav2vec27680.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typewav2vec2wav2vec2✓ kept
modelvocab_size3232✓ kept
modelhidden_actgelugelu✓ kept

Module classes

Instances of each class in the built model (prefix Wav2Vec2 omitted).

ClassOriginalFixture
Linear7414
Dropout399
LayerNorm266
GELUActivation2010
Attention122
EncoderLayer122
FeedForward122
Conv1d77
NoLayerNormConvLayer66
ModuleList22
GroupNorm11
ModuleDict11
ParametrizationList11
ParametrizedConv1d11
Encoder11
FeatureEncoder11
FeatureProjection11
ForCTC11
GroupNormConvLayer11
Model11
PositionalConvEmbedding11
SamePadLayer11
_WeightNorm11

Config changes

4 fields changed, 64 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/wav2vec2-mic.html b/shrink/wav2vec2-mic.html index fad6d26627ad41e1a9dcb3b71a6a398bdfa1ed39..f401b2437d33dfdcdf25e7bebcdad77aeff1b00c 100644 --- a/shrink/wav2vec2-mic.html +++ b/shrink/wav2vec2-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/wav2vec2-base-960h

huggingface-mic/wav2vec2-mic@4cce7c54a33f ← facebook/wav2vec2-base-960h@22aad52d435e · Wav2Vec2ForCTC · random weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 16 failed checks in 30 days · MIC failures dashboard

Original94.4Mparameters
Fixture4.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

wav2vec2.feature_extractor.conv_layers

Original · 7
Fixture · 7
KindOriginalFixture
GroupNorm, GroupNormConvLayer11✓ kept
NoLayerNormConvLayer66✓ kept

wav2vec2.encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
EncoderLayer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
wav2vec2.encoder.layers85.1M90.1%66.9K1.5%
wav2vec2.encoder4.7M5.0%33.1K0.8%
wav2vec2.feature_extractor.conv_layers4.2M4.4%4.2M96.9%
wav2vec2.feature_projection395.0K0.4%33.9K0.8%
lm_head24.6K0.0%2.1K0.0%
wav2vec27680.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typewav2vec2wav2vec2✓ kept
modelvocab_size3232✓ kept
modelhidden_actgelugelu✓ kept

Module classes

Instances of each class in the built model (prefix Wav2Vec2 omitted).

ClassOriginalFixture
Linear7414
Dropout399
LayerNorm266
GELUActivation2010
Attention122
EncoderLayer122
FeedForward122
Conv1d77
NoLayerNormConvLayer66
ModuleList22
GroupNorm11
ModuleDict11
ParametrizationList11
ParametrizedConv1d11
Encoder11
FeatureEncoder11
FeatureProjection11
ForCTC11
GroupNormConvLayer11
Model11
PositionalConvEmbedding11
SamePadLayer11
_WeightNorm11

Config changes

4 fields changed, 64 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/wavlm-base-plus-mic.html b/shrink/wavlm-base-plus-mic.html index 1154a82c30b79af9c8df13a190120b607b39ba2e..754a630ee299a08a071e6183415d8f22d9472f2f 100644 --- a/shrink/wavlm-base-plus-mic.html +++ b/shrink/wavlm-base-plus-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

microsoft/wavlm-base-plus

huggingface-mic/wavlm-base-plus-mic@55205db7dc5c ← microsoft/wavlm-base-plus@4c66d4806a42 · WavLMModel · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original94.4Mparameters
Fixture4.3Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

feature_extractor.conv_layers

Original · 7
Fixture · 7
KindOriginalFixture
GroupNorm, GroupNormConvLayer11✓ kept
NoLayerNormConvLayer66✓ kept

encoder.layers

Original · 12
Fixture · 2
KindOriginalFixture
Embedding11✓ kept
no Embedding111✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
encoder.layers85.1M90.1%68.5K1.6%
encoder.pos_conv_embed4.7M5.0%33.0K0.8%
feature_extractor.conv_layers4.2M4.5%4.2M96.9%
feature_projection.projection394.0K0.4%32.8K0.8%
encoder.layer_norm1.5K0.0%1280.0%
feature_projection.layer_norm1.0K0.0%1.0K0.0%
masked_spec_embed7680.0%640.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typewavlmwavlm✓ kept
modelvocab_size3232✓ kept
modelhidden_actgelugelu✓ kept

Module classes

Instances of each class in the built model (prefix WavLM omitted).

ClassOriginalFixture
Linear8515
Dropout388
LayerNorm266
GELUActivation2010
Attention122
EncoderLayer122
FeedForward122
Conv1d77
NoLayerNormConvLayer66
ModuleList22
Embedding11
GroupNorm11
ModuleDict11
ParametrizationList11
ParametrizedConv1d11
Encoder11
FeatureEncoder11
FeatureProjection11
GroupNormConvLayer11
Model11
PositionalConvEmbedding11
SamePadLayer11
_WeightNorm11

Config changes

4 fields changed, 76 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/whisper-mic.html b/shrink/whisper-mic.html index 2da58c531764a03899873b7c57124f7c3a73600d..6571f78b3a4ad7e65ada28ca001a619b2a1e6428 100644 --- a/shrink/whisper-mic.html +++ b/shrink/whisper-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

openai/whisper-tiny

huggingface-mic/whisper-mic@1f40ed2753f5 ← openai/whisper-tiny@169d4a4341b3 · WhisperForConditionalGeneration · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 7 failed checks in 30 days · MIC failures dashboard

Original37.8Mparameters
Fixture3.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 4
Fixture · 2
KindOriginalFixture
EncoderLayer42✓ kept

model.decoder.layers

Original · 4
Fixture · 2
KindOriginalFixture
DecoderLayer42✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.decoder20.1M53.2%3.3M92.0%
model.decoder.layers9.5M25.1%100.2K2.8%
model.encoder.layers7.1M18.8%66.8K1.8%
model.encoder1.1M2.9%123.9K3.4%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typewhisperwhisper✓ kept
modelvocab_size5186551865✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelnum_mel_bins8080✓ kept

Module classes

Instances of each class in the built model (prefix Whisper omitted).

ClassOriginalFixture
Linear6533
LayerNorm2212
Attention126
GELUActivation84
DecoderLayer42
EncoderLayer42
Conv1d22
Embedding22
ModuleList22
Decoder11
Encoder11
ForConditionalGeneration11
Model11
PositionalEmbedding11

Config changes

7 fields changed, 40 unchanged.

FieldOriginalFixture
d_model38464
decoder_attention_heads64
decoder_ffn_dim1536128
decoder_layers42
encoder_attention_heads64
encoder_ffn_dim1536128
encoder_layers42
\ No newline at end of file diff --git a/shrink/whisper-tiny-mic.html b/shrink/whisper-tiny-mic.html index f7875f4e4320d6da44d693fefd9ccdd6b1b02aa5..c20191ed71620937426841956eb1525a194955a4 100644 --- a/shrink/whisper-tiny-mic.html +++ b/shrink/whisper-tiny-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

openai/whisper-tiny

huggingface-mic/whisper-tiny-mic@9d85803a8fb0 ← openai/whisper-tiny@169d4a4341b3 · WhisperForConditionalGeneration · overfit weights · built before the shrink rule (no recorded recipe)

+

CI: passing now; 10 failed checks in 30 days · MIC failures dashboard

Original37.8Mparameters
Fixture3.6Mparameters
Layer stacks2compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

model.encoder.layers

Original · 4
Fixture · 2
KindOriginalFixture
EncoderLayer42✓ kept

model.decoder.layers

Original · 4
Fixture · 2
KindOriginalFixture
DecoderLayer42✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
model.decoder20.1M53.2%3.3M92.0%
model.decoder.layers9.5M25.1%100.2K2.8%
model.encoder.layers7.1M18.8%66.8K1.8%
model.encoder1.1M2.9%123.9K3.4%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typewhisperwhisper✓ kept
modelvocab_size5186551865✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelnum_mel_bins8080✓ kept

Module classes

Instances of each class in the built model (prefix Whisper omitted).

ClassOriginalFixture
Linear6533
LayerNorm2212
Attention126
GELUActivation84
DecoderLayer42
EncoderLayer42
Conv1d22
Embedding22
ModuleList22
Decoder11
Encoder11
ForConditionalGeneration11
Model11
PositionalEmbedding11

Config changes

7 fields changed, 40 unchanged.

FieldOriginalFixture
d_model38464
decoder_attention_heads64
decoder_ffn_dim1536128
decoder_layers42
encoder_attention_heads64
encoder_ffn_dim1536128
encoder_layers42
\ No newline at end of file diff --git a/shrink/xlm-roberta-base-mic.html b/shrink/xlm-roberta-base-mic.html index a50b09b89a0b393488352a6df6126cb0fabb86bd..14420b6cbdc95979cbf3072f970362adf1e48d3b 100644 --- a/shrink/xlm-roberta-base-mic.html +++ b/shrink/xlm-roberta-base-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

FacebookAI/xlm-roberta-base

huggingface-mic/xlm-roberta-base-mic@934d2300ab3a ← FacebookAI/xlm-roberta-base@e73636d4f797 · XLMRobertaForMaskedLM · random weights · built before the shrink rule (no recorded recipe)

+

CI: no failure on main in 30 days · MIC failures dashboard

Original278.3Mparameters
Fixture16.4Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

roberta.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
lm_head.decoder192.0M69.0%16.0M97.8%
roberta.encoder.layer85.1M30.6%66.9K0.4%
lm_head.dense590.6K0.2%4.2K0.0%
roberta.embeddings397.1K0.1%33.1K0.2%
lm_head250.0K0.1%250.0K1.5%
lm_head.layer_norm1.5K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typexlm-robertaxlm-roberta✓ kept
modelvocab_size250002250002✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings514514✓ kept

Module classes

Instances of each class in the built model (prefix XLMRoberta omitted).

ClassOriginalFixture
Linear7414
Dropout377
LayerNorm266
GELUActivation122
Attention122
Intermediate122
Layer122
Output122
SelfAttention122
SelfOutput122
Embedding33
ModuleList11
Embeddings11
Encoder11
ForMaskedLM11
LMHead11
Model11

Config changes

4 fields changed, 27 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122
\ No newline at end of file diff --git a/shrink/xlm-roberta-xl-mic.html b/shrink/xlm-roberta-xl-mic.html index cab9a8f34084235e3a32915ffc7370686b0e5736..d2c99390b6698d98651b535ec23dea95cafad84a 100644 --- a/shrink/xlm-roberta-xl-mic.html +++ b/shrink/xlm-roberta-xl-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

facebook/xlm-roberta-xl

huggingface-mic/xlm-roberta-xl-mic@54cda088a6d8 ← facebook/xlm-roberta-xl@aa5d12025584 · XLMRobertaXLForMaskedLM · random weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original3.5Bparameters
Fixture16.4Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

roberta.encoder.layer

Original · 36
Fixture · 2
KindOriginalFixture
Layer362✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
roberta.encoder.layer2.8B81.3%66.9K0.4%
roberta.embeddings643.6M18.5%16.1M98.0%
lm_head.dense6.6M0.2%4.2K0.0%
lm_head250.9K0.0%250.9K1.5%
lm_head.layer_norm5.1K0.0%1280.0%
roberta.encoder5.1K0.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typexlm-roberta-xlxlm-roberta-xl✓ kept
modelvocab_size250880250880✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelugelu✓ kept
modelmax_position_embeddings514514✓ kept

Module classes

Instances of each class in the built model (prefix XLMRobertaXL omitted).

ClassOriginalFixture
Linear21814
LayerNorm746
Dropout735
GELUActivation362
Attention362
Intermediate362
Layer362
Output362
SelfAttention362
SelfOutput362
Embedding33
ModuleList11
Embeddings11
Encoder11
ForMaskedLM11
LMHead11
Model11

Config changes

4 fields changed, 27 unchanged.

FieldOriginalFixture
hidden_size256064
intermediate_size10240128
num_attention_heads324
num_hidden_layers362
\ No newline at end of file diff --git a/shrink/yolos-small-mic.html b/shrink/yolos-small-mic.html index c77cf33df2cddb63e6de45190419e8831d7d61c6..85a9a1f0a8968f6c3a2357fd8c8ee12b11c8c950 100644 --- a/shrink/yolos-small-mic.html +++ b/shrink/yolos-small-mic.html @@ -32,5 +32,6 @@ tr.bad td{color:var(--warn)}

← MIC coverage of Transformers model types

hustvl/yolos-small

huggingface-mic/yolos-small-mic@908c43a6393a ← hustvl/yolos-small@3d8f7130d3ce · YolosForObjectDetection · overfit weights

+

CI: no failure on main in 30 days · MIC failures dashboard

Original30.7Mparameters
Fixture379.7Kparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

vit.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
vit.encoder.layer21.3M69.4%66.9K17.6%
vit.encoder7.7M25.2%117.1K30.8%
vit.embeddings1.0M3.4%172.7K45.5%
class_labels_classifier.layers331.1K1.1%14.3K3.8%
bbox_predictor.layers297.2K1.0%8.6K2.3%
vit.layernorm7680.0%1280.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typeyolosyolos✓ kept
modelhidden_actgelugelu✓ kept
modelimage_size[512, 864][512, 864]✓ kept
modelpatch_size1616✓ kept
modelnum_channels33✓ kept

Module classes

Instances of each class in the built model (prefix Yolos omitted).

ClassOriginalFixture
Linear7818
Dropout255
LayerNorm255
GELUActivation122
Attention122
Intermediate122
Layer122
Output122
SelfAttention122
SelfOutput122
ModuleList33
MLPPredictionHead22
Conv2d11
InterpolateInitialPositionEmbeddings11
InterpolateMidPositionEmbeddings11
Embeddings11
Encoder11
ForObjectDetection11
Model11
PatchEmbeddings11

Config changes

4 fields changed, 27 unchanged.

FieldOriginalFixture
hidden_size38464
intermediate_size1536128
num_attention_heads64
num_hidden_layers122
\ No newline at end of file diff --git a/tools/README.md b/tools/README.md index 9be2f3018d3b7a27b57911559c4636d36ec19025..3cec9e839701f5f5fa66c1e25138817c004d3371 100644 --- a/tools/README.md +++ b/tools/README.md @@ -14,6 +14,7 @@ Add `--rerender` after a change to tmic's `shrink_view`, so every page picks it - **Done / Reachable** come from `transformersci.contracts.coverage` (fleet contracts and the transformers-mic shrink sweep). - **No checkpoint** splits the sweep's "no usable checkpoint" rows using `experiments/results/sweep_sources.json` (`tools/nockpt.py`). - **Gated** checks the source checkpoint of each Done type, else its top checkpoint (`tools/gated.py`). +- **CI dot and shrink-page CI line:** one request to the dashboard's `/mic/data` (30 days, `main`) per refresh (`tools/results.py`); viewers never call the dashboard. A fixture with no entry had no failure in the window. - **MIC (shrink)** renders a page only for fixtures whose pinned revision changed (`tools/shrink.py`). `--python` must have a transformers that knows every fleet model type. diff --git a/tools/results.py b/tools/results.py new file mode 100644 index 0000000000000000000000000000000000000000..aa090a446de3de8486b6d0c56b7bf1ab254ea4e8 --- /dev/null +++ b/tools/results.py @@ -0,0 +1,32 @@ +"""MIC test results on Transformers main, from the CI dashboard: one request per refresh, none from viewers.""" +import json +import urllib.request +from urllib.parse import urlencode + +DASHBOARD = "https://transformers-ci.huggingface.cool" +FAILURES = f"{DASHBOARD}/d/mic-failures/mic-failures?from=now-4w&to=now&timezone=browser" +DAYS = 30 + + +def fixture_url(name): + """The dashboard's MIC page filtered to one fixture (the Grafana page does not forward filters).""" + return f"{DASHBOARD}/mic?" + urlencode({"q": name, "days": DAYS, "sticky_only": "false"}) + + +def fetch(): + """{fixture name: [failing now, failed checks in DAYS]} plus the dashboard summary. + + The dashboard lists only cases that failed in the window, so a fleet fixture with + no entry had no failure on main in DAYS days. + """ + query = urlencode({"days": DAYS, "sticky_only": "false", "scope": "main"}) + with urllib.request.urlopen(f"{DASHBOARD}/mic/data?{query}", timeout=120) as response: + data = json.load(response) + out = {} + for row in data["rows"]: + name = (row.get("fixture_repo") or "").split("/")[-1] + if not name: + continue + now, fails = out.setdefault(name, [0, 0]) + out[name] = [now + bool(row.get("sticky")), fails + (row.get("fails") or 0)] + return out, {"days": DAYS, "now": data["summary"]["now"], "dashboard": FAILURES} diff --git a/tools/shrink.py b/tools/shrink.py index b6fc5f5768ff48a0f8f39ecb92a8682c48f926d8..1df432674cefc6c7cb8ed186fca4caa190c40954 100644 --- a/tools/shrink.py +++ b/tools/shrink.py @@ -11,6 +11,8 @@ from huggingface_hub import hf_hub_download # Relative and same-frame: the Hub shows the Space in a sandboxed iframe that blocks target="_top". BACK = '

← MIC coverage of Transformers model types

\n' +CI_MARK = "" +CI_LINE = re.compile(r'

.*?

\n') OLD_BACK = re.compile(r'

]*>← MIC coverage[^<]*

\n') PAGE = """ @@ -47,15 +49,41 @@ def render(pins, pages, mic, python=sys.executable, workers=3): return [line for line in pool.map(one, pins) if line] -def link_back(pages): - """Put a link to the coverage page at the top of every shrink page.""" +def ci_line(name, outcome): + """One line on the fixture's results on Transformers main, linking the dashboard.""" + from results import DAYS, FAILURES, fixture_url + + now, fails = outcome or (0, 0) + state = ( + f"failing now on main ({fails} failed checks in {DAYS} days)" + if now + else f"passing now; {fails} failed checks in {DAYS} days" + if fails + else f"no failure on main in {DAYS} days" + ) + return ( + f'

{CI_MARK}CI: {state}' + f' · MIC failures dashboard

\n' + ) + + +def link_back(pages, outcomes=None): + """Put a link to the coverage page, and the fixture's CI results, at the top of every shrink page.""" + outcomes = outcomes or {} for page in pages.glob("*.html"): - text = OLD_BACK.sub("", page.read_text()) - page.write_text(text.replace("
\n

", f"
\n{BACK}

", 1)) + text = CI_LINE.sub("", OLD_BACK.sub("", page.read_text())) + text = text.replace("
\n

", f"
\n{BACK}

", 1) + meta_end = text.index("

\n", text.index("

")) + len("

\n") # the pins line under the title + page.write_text(text[:meta_end] + ci_line(page.stem, outcomes.get(page.stem)) + text[meta_end:]) + +def links(fixtures, pages, fleets=None): + """fixtures: (repo, revision) pins -> {model_type: [[page name, findings, fleet, source], ...]} -def links(fixtures, pages): - """fixtures: (repo, revision) pins -> {model_type: [[page name, findings], ...]}""" + fleets: {repo: "control" | "prototypes"}; a model type can have a hand-written control + contract and a tmic prototype, or fixtures built from different checkpoints. + """ + fleets = fleets or {} out = {} for repo, revision in fixtures: name = repo.split("/")[-1] @@ -64,7 +92,8 @@ def links(fixtures, pages): continue findings = len(json.loads(record.read_text()).get("findings", [])) if record.exists() else 0 model_type = json.loads(Path(hf_hub_download(repo, "config.json", revision=revision)).read_text())["model_type"] - entry = [name, findings] + build = json.loads(Path(hf_hub_download(repo, "build.json", revision=revision)).read_text()) + entry = [name, findings, fleets.get(repo, ""), (build.get("built_from") or {}).get("repo", "")] if entry not in out.setdefault(model_type, []): out[model_type].append(entry) return {k: sorted(v) for k, v in sorted(out.items())} diff --git a/tools/update.py b/tools/update.py index 2c67ea6dbc87646a74788d91c62f3c4b3d2e08a5..e98ccdf4c220fd2a8045d34b30bea497a2c0ddf6 100644 --- a/tools/update.py +++ b/tools/update.py @@ -37,6 +37,7 @@ from transformersci.contracts import coverage # noqa: E402 import gated # noqa: E402 import nockpt # noqa: E402 +import results # noqa: E402 import shrink # noqa: E402 cov = coverage.build() @@ -55,8 +56,14 @@ todo = [f"{repo}@{revision}" for repo, revision in fixtures] if args.rerender el print(f"rendering {len(todo)} shrink page(s)") for error in shrink.render(todo, pages, args.mic, args.python): print(" failed", error) -shrink.link_back(pages) -data["shrink"] = shrink.links(fixtures, pages) +data["results"], data["meta"]["results"] = results.fetch() +shrink.link_back(pages, data["results"]) +fleets = {} +for name in ("prototypes", "control"): # control wins for a repo in both + base = f"https://huggingface.co/datasets/{coverage.fleet.REPOSITORY}/resolve/{cov['fleet']['revision']}" + for row in coverage.fleet._rows("contracts", coverage._get(f"{base}/{name}/contracts.csv")): + fleets[row["mic_repo"]] = name +data["shrink"] = shrink.links(fixtures, pages, fleets) data["gated"] = gated.gating(data["rows"], fixtures) data["meta"].update( generated=datetime.date.today().isoformat(), fleet=cov["fleet"]["revision"], sweep=cov["sweep"]["revision"]