File size: 5,385 Bytes
c2e902a 6466ca1 c2e902a 6466ca1 c2e902a 5d84e23 c2e902a 5d84e23 c2e902a 6466ca1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 | import json
from pathlib import Path
import pytest
from safetensors.torch import save_file
from MVP.qwen35_prune import (
ParameterReport,
build_text_config,
count_parameter_groups,
choose_prefix,
translate_text_key,
)
from MVP.validate_checkpoint import validate_state_dict_keys
FIXTURE = Path(__file__).parent / "fixtures" / "qwen35_metadata.json"
def load_fixture():
return json.loads(FIXTURE.read_text())
def test_measured_n4_prefix_is_inside_required_interval():
data = load_fixture()
report = ParameterReport(
embedding_params=data["embedding_params"],
layer_params=tuple(
[data["linear_attention_params"]] * 3
+ [data["full_attention_params"]]
+ [data["linear_attention_params"]] * 3
+ [data["full_attention_params"]] * 5
),
layer_types=tuple(data["layer_types"]),
final_norm_params=data["final_norm_params"],
all_named_params=data["all_named_params"],
)
choice = choose_prefix(report, 330_000_000, 350_000_000)
assert choice.layer_count == 4
assert choice.parameter_count == 337_299_424
assert 330_000_000 <= choice.parameter_count <= 350_000_000
def test_prefix_selection_rejects_incomplete_hybrid_block():
data = load_fixture()
report = ParameterReport(
embedding_params=data["embedding_params"],
layer_params=(data["linear_attention_params"],) * 24,
layer_types=tuple(data["layer_types"]),
final_norm_params=data["final_norm_params"],
all_named_params=data["all_named_params"],
)
with pytest.raises(ValueError, match="complete hybrid"):
choose_prefix(report, 330_000_000, 350_000_000, requested_layers=3)
def test_text_prefix_translation_drops_vision_and_mtp():
assert (
translate_text_key("model.language_model.layers.3.linear_attn.A_log")
== "model.layers.3.linear_attn.A_log"
)
assert translate_text_key("model.visual.patch_embed.proj.weight") is None
assert translate_text_key("mtp.layers.0.mlp.down_proj.weight") is None
def test_text_config_is_standalone_and_keeps_live_attention_fields():
full = {
"model_type": "qwen3_5",
"tie_word_embeddings": True,
"vision_config": {"hidden_size": 512},
"text_config": {
"model_type": "qwen3_5_text",
"hidden_size": 1024,
"intermediate_size": 3584,
"num_hidden_layers": 24,
"layer_types": list(load_fixture()["layer_types"]),
"linear_num_key_heads": 16,
"linear_num_value_heads": 16,
"linear_key_head_dim": 128,
"linear_value_head_dim": 128,
"linear_conv_kernel_dim": 4,
"vocab_size": 248320,
"mtp_num_hidden_layers": 1,
"mtp_use_dedicated_embeddings": False,
},
}
text = build_text_config(full, 4)
assert text["model_type"] == "qwen3_5_text"
assert text["num_hidden_layers"] == 4
assert text["layer_types"] == load_fixture()["layer_types"][:4]
assert text["linear_num_value_heads"] == 16
assert text["tie_word_embeddings"] is True
assert "vision_config" not in text
assert not any(key.startswith("mtp_") for key in text)
def test_validator_rejects_vision_mtp_and_duplicate_tied_head():
config = {
"model_type": "qwen3_5_text",
"tie_word_embeddings": True,
"num_hidden_layers": 4,
"layer_types": load_fixture()["layer_types"][:4],
}
keys = {
"model.embed_tokens.weight": (248320, 1024),
"model.layers.0.input_layernorm.weight": (1024,),
"model.layers.1.input_layernorm.weight": (1024,),
"model.layers.2.input_layernorm.weight": (1024,),
"model.layers.3.input_layernorm.weight": (1024,),
"model.norm.weight": (1024,),
"lm_head.weight": (248320, 1024),
"model.visual.patch_embed.proj.weight": (1, 1),
"mtp.fc.weight": (1, 1),
}
with pytest.raises(ValueError, match="vision|MTP|tied"):
validate_state_dict_keys(keys, config, 330_000_000, 350_000_000)
def test_metadata_counter_uses_tensor_shapes_without_loading_model():
data = load_fixture()
config = {
"text_config": {
"layer_types": list(data["layer_types"][:4]),
}
}
tensors = {
"model.language_model.embed_tokens.weight": (2, 3),
"model.language_model.layers.0.input_layernorm.weight": (2,),
"model.language_model.layers.1.input_layernorm.weight": (3,),
"model.language_model.layers.2.input_layernorm.weight": (4,),
"model.language_model.layers.3.input_layernorm.weight": (5,),
"model.language_model.norm.weight": (2,),
"model.visual.patch_embed.proj.weight": (7,),
"mtp.fc.weight": (11,),
}
def make_tensor(shape):
import torch
return torch.zeros(shape, dtype=torch.float32)
path = FIXTURE.parent / "counter-fixture.safetensors"
try:
save_file({key: make_tensor(shape) for key, shape in tensors.items()}, str(path))
report = count_parameter_groups(path, config)
finally:
path.unlink(missing_ok=True)
assert report.embedding_params == 6
assert report.layer_params == (2, 3, 4, 5)
assert report.final_norm_params == 2
assert report.all_named_params == 40
|