---
library_name: transformers
base_model:
- zai-org/GLM-5.3-Flash
---
This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash).
Structural comparison with `zai-org/GLM-5.3-Flash`:
| Feature | Source | Tiny |
|---|---:|---:|
| Checkpoint tensor elements | 321.32B | 112.09M, including 16.70K FP32 scale elements |
| Safetensors weight files | 328.34 GB across shards | 207.31 MB (`model.safetensors`) |
| Architecture / modalities | GLM5-Next; text, image, video | Same |
| Vocabulary rows / tied embeddings | 154,880 / false | Same |
| Text layers / attention cadence | 45; three KDA then DSA | 8; two complete three-KDA + DSA cycles |
| Text hidden size | 4,096 | 128 |
| KDA heads / head dimension | 64 / 128 | 64 / 128 |
| DSA heads / NoPE / value dimensions | 64 / 256 / 256 | 64 / 32 / 256 |
| Indexer heads / dimension / top-k | 32 / 128 / 2,048 | 32 / 128 / 128 |
| Routed experts / selected experts / shared experts | 288 / 8 / 1 | 32 / 8 / 1 |
| Hyper-Connections | Four manifold-constrained streams | Same |
| MTP | One DSA/MoE layer with shared embedding/head | Same |
| Vision depth / hidden / output size | 24 / 1,024 / 4,096 | 2 / 128 / 128 |
| Checkpoint precision | BF16 plus E4M3 block-FP8 and FP32 scales | Same; 32x32 blocks for reduced axes |
The KDA and DSA indexer heads remain 128-wide because vLLM's GLM5-Next KDA and Hadamard-128 indexer kernels require that ABI. The source E4M3 format is retained with reduced 32x32 blocks; the TP=4 example selects vLLM's Triton block-FP8 GEMM because H20's automatic Humming kernel requires an output scale group of at least 64. With the unchanged 32-wide FP8 K block, vLLM TP=4 partitions residual-width linear inputs by four. The 32-wide candidate reached GLM5-Next model loading but was rejected with `input_size_per_partition = 8 is not divisible by weight quantization block_k = 32`; a 64-wide residual would likewise partition to 16. The tested 128-wide checkpoint is therefore the smallest executable residual width for this required checkpoint/backend/TP combination.
Tokenizer, processor, chat template and generation assets are copied unchanged from the pinned source revision.
| File path | Size |
|------|------|
| model.safetensors | 207.3MB |
### Example usage:
- vLLM (✅ verified)
```bash
vllm serve tiny-random/glm-5.3-flash \
--tensor-parallel-size 4 \
--speculative-config '{"method":"mtp","num_speculative_tokens":1}' \
--max-model-len 4096 \
--max-num-batched-tokens 512 \
--max-num-seqs 4 \
--gpu-memory-utilization 0.3 \
--kernel-config '{"moe_backend":"triton","linear_backend":"triton"}' \
--compilation-config '{"cudagraph_capture_sizes":[1,2,4,8]}'
```
- Transformers (✅ verified)
```python
import numpy as np
import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor
from transformers.integrations.hub_kernels import allow_all_hub_kernels
from transformers.video_utils import VideoMetadata
model_id = "tiny-random/glm-5.3-flash"
# Transformers resolves the official fine-grained FP8 kernel from the Hub.
with allow_all_hub_kernels():
model = AutoModelForImageTextToText.from_pretrained(
model_id, dtype=torch.bfloat16, attn_implementation="sdpa",
).to("cuda")
processor = AutoProcessor.from_pretrained(model_id)
image = Image.new("RGB", (28, 28), "red")
video = np.stack([
np.asarray(Image.new("RGB", (28, 28), color))
for color in ("red", "green", "blue", "yellow")
])
messages = [{"role": "user", "content": [
{"type": "image", "image": image},
{"type": "video", "video": video},
{"type": "text", "text": "Describe the image and the video."},
]}]
inputs = processor.apply_chat_template(
messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",
video_metadata=[VideoMetadata(total_num_frames=4, fps=2, frames_indices=[0, 1, 2, 3])],
).to(model.device)
output = model.generate(**inputs, do_sample=False, max_new_tokens=10)
print(processor.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```
### Codes to create this repo (✅ verified):
Click to expand
```python
"""Create the GLM5-Next multimodal checkpoint in its public FP8 schema."""
import json
from pathlib import Path
from shutil import copyfile
import torch
from huggingface_hub import hf_hub_download, list_repo_files
from safetensors.torch import save_file
from transformers import AutoConfig, AutoModelForImageTextToText, set_seed
source_model_id = "zai-org/GLM-5.3-Flash"
save_folder = "/tmp/tiny-random/glm-53-flash"
source_revision = "46a050bec13002ad29ee795c16e61cf4fa3c4a9c"
set_seed(42)
folder = Path(save_folder)
folder.mkdir(parents=True, exist_ok=True)
raw = json.loads(Path(hf_hub_download(
source_model_id, "config.json", revision=source_revision)).read_text())
raw["quantization_config"]["weight_block_size"] = [32, 32]
text = raw["text_config"]
text.update(
hidden_size=128,
num_hidden_layers=8,
# Retain the source's TP=4-friendly global head topology. The KDA
# and DSA runtime kernels operate on 16 heads per TP rank.
num_attention_heads=64,
num_key_value_heads=64,
intermediate_size=128,
moe_intermediate_size=128,
n_routed_experts=32,
num_experts_per_tok=8,
q_lora_rank=128,
# FA3 sparse-MQA uses this latent as q_v and requires >=256.
kv_lora_rank=256,
qk_nope_head_dim=32,
qk_rope_head_dim=0,
# Retain source value-head width required by sparse-MQA FA3.
v_head_dim=256,
index_head_dim=128,
# The Hadamard-128 indexer fused kernel accepts 32 or 64; source is 32.
index_n_heads=32,
index_topk=128,
index_kpool=4,
linear_head_dim=128,
first_k_dense_replace=3,
layer_types=["linear_attention", "linear_attention", "linear_attention", "indexed_attention"] * 2,
mlp_layer_types=["dense", "dense", "dense", "sparse", "sparse", "sparse", "sparse", "sparse"],
indexer_types=["full"] * 8,
)
text["linear_attn_config"].update(
head_dim=128,
num_heads=64,
kda_layers=[0, 1, 2, 4, 5, 6],
full_attn_layers=[3, 7],
)
# TP=4 shards vision attention too: retain four visual heads and a
# 32-wide visual head dimension for an executable multimodal tower.
raw["vision_config"].update(
depth=2, hidden_size=128, intermediate_size=256, num_heads=4,
out_hidden_size=128, projection_intermediate_size=64, image_size=28, patch_size=14,
)
config_for_build = dict(raw)
config_for_build.pop("quantization_config")
(folder / "config.json").write_text(json.dumps(config_for_build, indent=2) + "\n")
config = AutoConfig.from_pretrained(folder)
model = AutoModelForImageTextToText.from_config(config, dtype=torch.bfloat16)
with torch.no_grad():
for name, value in model.named_parameters():
if name.endswith("A_log"):
value.zero_()
elif name.endswith("dt_bias"):
value.fill_(1)
elif name.endswith("e_score_correction_bias"):
value.zero_()
elif name.endswith("weight") and "norm" in name:
value.fill_(1)
else:
value.normal_(mean=0.0, std=0.02)
state = {name: value.detach().cpu().contiguous() for name, value in model.state_dict().items()}
# Transformers builds the published GLM5-Next graph with convenient
# module names (``attn_hc`` / ``forget_gate`` / fused ``conv1d``), while
# both the pinned source checkpoint and vLLM consume the public checkpoint
# schema below. Convert it before quantization: this is a lossless key
# rewrite except for splitting the concatenated Q|K|V convolution weight.
for name in list(state):
replacement = (
name.replace(".attn_hc.base", ".hc_attn_base")
.replace(".attn_hc.fn", ".hc_attn_fn")
.replace(".attn_hc.scale", ".hc_attn_scale")
.replace(".ffn_hc.base", ".hc_ffn_base")
.replace(".ffn_hc.fn", ".hc_ffn_fn")
.replace(".ffn_hc.scale", ".hc_ffn_scale")
.replace(".self_attn.forget_gate.", ".self_attn.")
)
if replacement != name:
state[replacement] = state.pop(name)
projection_size = text["linear_attn_config"]["num_heads"] * text["linear_attn_config"]["head_dim"]
for name in list(state):
if not name.endswith(".self_attn.conv1d.weight"):
continue
prefix = name.removesuffix(".conv1d.weight")
q, k, v = state.pop(name).split(projection_size, dim=0)
state[prefix + ".q_conv1d.weight"] = q.contiguous()
state[prefix + ".k_conv1d.weight"] = k.contiguous()
state[prefix + ".v_conv1d.weight"] = v.contiguous()
block_out, block_in = raw["quantization_config"]["weight_block_size"]
def pack(name: str, weight: torch.Tensor) -> None:
rows, columns = weight.shape
assert rows % block_out == 0 and columns % block_in == 0, (name, weight.shape)
blocks = weight.float().reshape(rows // block_out, block_out, columns // block_in, block_in)
scales = (blocks.abs().amax(dim=(1, 3)) / 448).clamp_min(1e-12).to(torch.float32)
state[name] = (blocks / scales[:, None, :, None]).clamp(-448, 448).reshape(weight.shape).to(
torch.float8_e4m3fn
).contiguous()
state[name.removesuffix("weight") + "weight_scale_inv"] = scales.contiguous()
def keep_bf16(name: str) -> bool:
# GLM5-Next keeps every KDA/linear-attention tensor in BF16. They are
# exposed below ``self_attn`` in the Transformers graph rather than
# under a ``linear_attention`` component name.
kda_prefixes = tuple(
f"model.language_model.layers.{index}.self_attn."
for index in (0, 1, 2, 4, 5, 6)
)
return (
name in {"model.language_model.embed_tokens.weight", "lm_head.weight"}
or name.startswith(kda_prefixes)
# GLM5-Next's DSA value expansion deliberately remains BF16.
or name.endswith(".self_attn.kv_b_proj.weight")
or ".norm." in name
or ".indexer." in name
or ".linear_attention." in name
or name.endswith(".mlp.gate.weight")
or name.endswith(".mlp.gate.e_score_correction_bias")
or name.startswith("model.visual.")
)
# The upstream skip list contains absolute source-layer paths. Retain
# only the reduced graph's BF16 components so the Transformers FP8 loader
# constructs KDA, DSA value, router, indexer and vision modules correctly.
exclusions = {"model.language_model.embed_tokens", "lm_head", "model.visual"}
for module_name, module in model.named_modules():
if isinstance(module, torch.nn.Linear) and keep_bf16(module_name + ".weight"):
exclusions.add(module_name)
raw["quantization_config"]["modules_to_not_convert"] = sorted(exclusions)
for name in list(state):
if not name.endswith(".mlp.experts.gate_up_proj"):
continue
prefix = name.removesuffix(".gate_up_proj")
gate_up = state.pop(name)
down = state.pop(prefix + ".down_proj")
for expert in range(gate_up.shape[0]):
gate, up = gate_up[expert].chunk(2, dim=0)
pack(f"{prefix}.{expert}.gate_proj.weight", gate)
pack(f"{prefix}.{expert}.up_proj.weight", up)
pack(f"{prefix}.{expert}.down_proj.weight", down[expert])
for name in list(state):
value = state[name]
if not (name.endswith(".weight") and value.ndim == 2) or keep_bf16(name):
continue
pack(name, value)
# vLLM consumes the next-token predictor from layer index num_hidden_layers.
# It is a DSA/MoE layer, not a KDA/mHC layer, and shares the global token
# embedding and LM head exactly like the source checkpoint.
base_prefix = "model.language_model.layers.7."
mtp_prefix = "model.language_model.layers.8."
for name, value in list(state.items()):
# GLM5-Next's embedded draft DSA/MoE layer has no Hyper-Connections;
# source layer 45 likewise omits every hc_* tensor.
if name.startswith(base_prefix) and ".hc_" not in name:
state[mtp_prefix + name.removeprefix(base_prefix)] = value.clone()
state[mtp_prefix + "eh_proj.weight"] = torch.empty(128, 256, dtype=torch.bfloat16).normal_(0, 0.02)
for name in ("enorm.weight", "hnorm.weight", "shared_head.norm.weight"):
state[mtp_prefix + name] = torch.ones(128, dtype=torch.bfloat16)
save_file(state, folder / "model.safetensors", metadata={"format": "pt"})
(folder / "config.json").write_text(json.dumps(raw, indent=2) + "\n")
assets = {
"tokenizer.json", "tokenizer_config.json", "special_tokens_map.json", "added_tokens.json",
"chat_template.jinja", "chat_template.json", "processor_config.json", "preprocessor_config.json",
"video_preprocessor_config.json", "generation_config.json", "LICENSE",
}
for name in sorted(set(list_repo_files(source_model_id, revision=source_revision)) & assets):
copyfile(hf_hub_download(source_model_id, name, revision=source_revision), folder / name)
```
Click to expand
```text
Glm5NextForConditionalGeneration(
(model): Glm5NextModel(
(visual): Glm5NextVisionModel(
(patch_embed): Glm5NextVisionPatchEmbed(
(proj): Conv3d(3, 128, kernel_size=(2, 14, 14), stride=(2, 14, 14))
)
(rotary_pos_emb): Glm5NextVisionRotaryEmbedding()
(blocks): ModuleList(
(0-1): 2 x Glm5NextVisionBlock(
(norm1): Glm5NextRMSNorm((128,), eps=1e-05)
(norm2): Glm5NextRMSNorm((128,), eps=1e-05)
(attn): Glm5NextVisionAttention(
(qkv): Linear(in_features=128, out_features=384, bias=True)
(proj): Linear(in_features=128, out_features=128, bias=True)
(q_norm): Glm5NextRMSNorm((32,), eps=1e-05)
(k_norm): Glm5NextRMSNorm((32,), eps=1e-05)
)
(mlp): Glm5NextVisionMLP(
(gate_proj): Linear(in_features=128, out_features=256, bias=True)
(up_proj): Linear(in_features=128, out_features=256, bias=True)
(down_proj): Linear(in_features=256, out_features=128, bias=True)
(act_fn): SiLUActivation()
)
)
)
(merger): Glm5NextVisionPatchMerger(
(proj): Linear(in_features=128, out_features=128, bias=False)
(post_projection_norm): LayerNorm((128,), eps=1e-05, elementwise_affine=True, bias=True)
(gate_proj): Linear(in_features=128, out_features=64, bias=False)
(up_proj): Linear(in_features=128, out_features=64, bias=False)
(down_proj): Linear(in_features=64, out_features=128, bias=False)
(act1): GELU(approximate='none')
(act_fn): SiLUActivation()
)
(downsample): Conv2d(128, 128, kernel_size=(2, 2), stride=(2, 2))
(post_layernorm): Glm5NextRMSNorm((128,), eps=1e-05)
)
(language_model): Glm5NextTextModel(
(embed_tokens): Embedding(154880, 128, padding_idx=154820)
(layers): ModuleList(
(0-2): 3 x Glm5NextTextDecoderLayer(
(self_attn): Glm5NextTextLinearAttention(
(q_proj): Linear(in_features=128, out_features=8192, bias=False)
(k_proj): Linear(in_features=128, out_features=8192, bias=False)
(v_proj): Linear(in_features=128, out_features=8192, bias=False)
(conv1d): Conv1d(24576, 24576, kernel_size=(4,), stride=(1,), padding=(3,), groups=24576, bias=False)
(forget_gate): Glm5NextTextForgetGate(
(f_a_proj): Linear(in_features=128, out_features=128, bias=False)
(f_b_proj): Linear(in_features=128, out_features=8192, bias=False)
)
(b_proj): Linear(in_features=128, out_features=64, bias=False)
(g_a_proj): Linear(in_features=128, out_features=128, bias=False)
(g_b_proj): Linear(in_features=128, out_features=8192, bias=False)
(o_norm): Glm5NextTextRMSNormGated()
(o_proj): Linear(in_features=8192, out_features=128, bias=False)
)
(mlp): Glm5NextTextMLP(
(gate_proj): Linear(in_features=128, out_features=128, bias=False)
(up_proj): Linear(in_features=128, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=128, bias=False)
(act_fn): SiLUActivation()
)
(input_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(post_attention_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(attn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
(ffn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
)
(3): Glm5NextTextDecoderLayer(
(self_attn): Glm5NextTextAttention(
(q_a_proj): Linear(in_features=128, out_features=128, bias=False)
(q_a_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(q_b_proj): Linear(in_features=128, out_features=2048, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=128, out_features=256, bias=False)
(kv_a_layernorm): Glm5NextTextRMSNorm((256,), eps=1e-05)
(kv_b_proj): Linear(in_features=256, out_features=18432, bias=False)
(o_proj): Linear(in_features=16384, out_features=128, bias=False)
(indexer): Glm5NextTextIndexer(
(wq_b): Linear(in_features=128, out_features=4096, bias=False)
(wk): Linear(in_features=128, out_features=128, bias=False)
(k_norm): LayerNorm((128,), eps=1e-06, elementwise_affine=True, bias=True)
(weights_proj): Linear(in_features=128, out_features=32, bias=False)
)
)
(mlp): Glm5NextTextMoE(
(experts): Glm5NextTextExperts()
(gate): Glm5NextTextTopkRouter()
(shared_experts): Glm5NextTextMLP(
(gate_proj): Linear(in_features=128, out_features=128, bias=False)
(up_proj): Linear(in_features=128, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=128, bias=False)
(act_fn): SiLUActivation()
)
)
(input_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(post_attention_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(attn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
(ffn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
)
(4-6): 3 x Glm5NextTextDecoderLayer(
(self_attn): Glm5NextTextLinearAttention(
(q_proj): Linear(in_features=128, out_features=8192, bias=False)
(k_proj): Linear(in_features=128, out_features=8192, bias=False)
(v_proj): Linear(in_features=128, out_features=8192, bias=False)
(conv1d): Conv1d(24576, 24576, kernel_size=(4,), stride=(1,), padding=(3,), groups=24576, bias=False)
(forget_gate): Glm5NextTextForgetGate(
(f_a_proj): Linear(in_features=128, out_features=128, bias=False)
(f_b_proj): Linear(in_features=128, out_features=8192, bias=False)
)
(b_proj): Linear(in_features=128, out_features=64, bias=False)
(g_a_proj): Linear(in_features=128, out_features=128, bias=False)
(g_b_proj): Linear(in_features=128, out_features=8192, bias=False)
(o_norm): Glm5NextTextRMSNormGated()
(o_proj): Linear(in_features=8192, out_features=128, bias=False)
)
(mlp): Glm5NextTextMoE(
(experts): Glm5NextTextExperts()
(gate): Glm5NextTextTopkRouter()
(shared_experts): Glm5NextTextMLP(
(gate_proj): Linear(in_features=128, out_features=128, bias=False)
(up_proj): Linear(in_features=128, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=128, bias=False)
(act_fn): SiLUActivation()
)
)
(input_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(post_attention_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(attn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
(ffn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
)
(7): Glm5NextTextDecoderLayer(
(self_attn): Glm5NextTextAttention(
(q_a_proj): Linear(in_features=128, out_features=128, bias=False)
(q_a_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(q_b_proj): Linear(in_features=128, out_features=2048, bias=False)
(kv_a_proj_with_mqa): Linear(in_features=128, out_features=256, bias=False)
(kv_a_layernorm): Glm5NextTextRMSNorm((256,), eps=1e-05)
(kv_b_proj): Linear(in_features=256, out_features=18432, bias=False)
(o_proj): Linear(in_features=16384, out_features=128, bias=False)
(indexer): Glm5NextTextIndexer(
(wq_b): Linear(in_features=128, out_features=4096, bias=False)
(wk): Linear(in_features=128, out_features=128, bias=False)
(k_norm): LayerNorm((128,), eps=1e-06, elementwise_affine=True, bias=True)
(weights_proj): Linear(in_features=128, out_features=32, bias=False)
)
)
(mlp): Glm5NextTextMoE(
(experts): Glm5NextTextExperts()
(gate): Glm5NextTextTopkRouter()
(shared_experts): Glm5NextTextMLP(
(gate_proj): Linear(in_features=128, out_features=128, bias=False)
(up_proj): Linear(in_features=128, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=128, bias=False)
(act_fn): SiLUActivation()
)
)
(input_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(post_attention_layernorm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(attn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
(ffn_hc): Glm5NextTextHyperConnection(
(input_norm): Glm5NextTextUnweightedRMSNorm()
)
)
)
(norm): Glm5NextTextRMSNorm((128,), eps=1e-05)
(hc_head): Glm5NextTextHyperHead()
)
)
(lm_head): Linear(in_features=128, out_features=154880, bias=False)
)
```