File size: 6,272 Bytes
1b8b8cd 0434c2c 1b8b8cd 0434c2c 1b8b8cd | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 | ---
library_name: transformers
pipeline_tag: image-text-to-text
base_model:
- mistralai/Mistral-Large-4-1T-A52B
---
This is a randomly initialized tiny model based on [mistralai/Mistral-Large-4-1T-A52B](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B) for testing and debugging.
| Feature | Source | This repo |
|---|---|---|
| Checkpoint tensor elements | 7.13B BF16, 63.84M F32, 1.05T F8_E4M3 | 54.84M BF16, 320 F32, 1.31M F8_E4M3 |
| Safetensors weight files | 1.06 TB (54 files) | 111.13 MB (consolidated.safetensors) |
| Architecture / input | Multimodal MoE; text and image | Same |
| Vocabulary rows / tied embeddings | 204,800 / false | 204,800 / true |
| Backbone layers | 72 | 12: three complete RoPE/RoPE/RoPE/NoPE cycles |
| Text hidden size / attention head dimension | 6,144 / 64 | 128 / 32 |
| Routed / selected / shared experts | 128 / 4 / 1 | 32 / 4 / 1 |
| DSA indexer | 32 heads; 128-wide; top-2,048 | 4 heads; 128-wide; top-128 |
| Cross-layer KV sharing | Source pattern | Two complete three-own/one-shared NoPE cycles |
| Vision tower | 30 layers; 2,048 width | 4 layers; 64 width |
| DSpark | 3 dense MLA stages; rank 512; depth 3 | 3 dense MLA stages; rank 8; depth 3 |
| Checkpoint precision | E4M3 block-FP8 + FP32 scales; BF16 exclusions | Same format; 64×64 blocks |
TP=2 vLLM uses a 64-wide FP8 activation group per rank, which fixes the residual width at 128. The checkpoint retains 64×64 FP8 blocks, a 512-wide MLA latent KV required by H20 sparse MLA, and four 32-wide attention heads. The input/output BF16 token table is tied, saving one 52.43 MB matrix; the optimized vLLM kernel constraints still place the checkpoint above the 30 MB soft target.
| File path | Size |
|------|------|
| consolidated.safetensors | 111.1MB |
### Example usage
#### Transformers (✅ verified)
```python
def run_transformers(model_id: str):
import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor
from transformers.utils.quantization_config import FineGrainedFP8Config
model = AutoModelForImageTextToText.from_pretrained(
str(model_id),
dtype=torch.bfloat16,
device_map="cuda",
attn_implementation="sdpa",
quantization_config=FineGrainedFP8Config(dequantize=True),
)
processor = AutoProcessor.from_pretrained(str(model_id))
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": Image.new("RGB", (56, 56), (30, 120, 220))},
{"type": "text", "text": "Describe this image in one word."},
],
}
]
inputs = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
reasoning_effort="low",
).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))
if __name__ == "__main__":
run_transformers(model_id="tiny-random/mistral-large-4")
```
#### vLLM offline (✅ verified)
Install the snapshot package with `pip install --no-deps <model directory>` before starting vLLM.
```python
def run_vllm(model_id: str):
import os
os.environ.setdefault("VLLM_USE_FLASHINFER_SAMPLER", "0")
os.environ.setdefault("VLLM_ALLREDUCE_USE_FLASHINFER", "0")
os.environ.setdefault("VLLM_ALLREDUCE_USE_SYMM_MEM", "0")
from PIL import Image
from transformers import AutoProcessor
from vllm import LLM, SamplingParams
model_id = str(model_id)
processor = AutoProcessor.from_pretrained(model_id)
llm = LLM(
model=model_id,
trust_remote_code=True,
tensor_parallel_size=2,
speculative_config={"method": "dspark", "num_speculative_tokens": 3},
kv_cache_dtype="bfloat16",
max_model_len=512,
max_num_batched_tokens=1024,
max_num_seqs=2,
gpu_memory_utilization=0.1,
disable_custom_all_reduce=True,
kernel_config={"linear_backend": "triton"},
attention_config={"backend": "FLASHINFER_MLA_SPARSE_SM90"},
)
sampling = SamplingParams(temperature=0, max_tokens=10, ignore_eos=True)
try:
print(llm.generate(["Hello"], sampling)[0].outputs[0].text)
image = Image.new("RGB", (56, 56), "blue")
inputs = processor.apply_chat_template(
[
{
"role": "user",
"content": [
{"type": "image", "image": image},
{"type": "text", "text": "Describe this image in one word."},
],
}
],
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
reasoning_effort="low",
)
result = llm.generate(
{"prompt_token_ids": inputs["input_ids"][0].tolist(), "multi_modal_data": {"image": image}},
sampling,
)
print(result[0].outputs[0].text)
finally:
llm.llm_engine.engine_core.shutdown(timeout=30)
if __name__ == "__main__":
run_vllm(model_id="tiny-random/mistral-large-4")
```
#### vLLM serving (unverified)
Use the official Mistral Large 4 vLLM image.
```bash
docker run --rm --runtime nvidia --gpus all --ipc=host -p 8000:8000 --entrypoint bash \
-v "$(pwd)/mistral-large-4:/model:ro" vllm/vllm-openai:mistral-large-4 \
-lc 'pip install --no-deps /model && vllm serve /model \
--tensor-parallel-size 2 \
--speculative-config '{"method":"dspark","num_speculative_tokens":3}' \
--kv-cache-dtype bfloat16 --disable-custom-all-reduce \
--kernel-config '{"linear_backend":"triton"}' \
--attention-config '{"backend":"FLASHINFER_MLA_SPARSE_SM90"}' \
--limit-mm-per-prompt '{"image":1}' --max-model-len 2048 \
--max-num-batched-tokens 256 --max-num-seqs 4 --gpu-memory-utilization 0.2'
```
### Creating this repo
- Code: [.meta/create_model.py](.meta/create_model.py)
- Model structure: [.meta/model.log](.meta/model.log)
- Environment: [.meta/env.json](.meta/env.json)
*This repo is generated by GPT-6.1-Sol.* |