gpt-oss-mxfp4

This tiny model is for debugging. It is randomly initialized with the config adapted from openai/gpt-oss-120b.

Note: This model used quantized MXFP4 FFN. pip install -U triton git+https://github.com/triton-lang/triton.git@main#subdirectory=python/triton_kernels

Example usage:

• vLLM

vllm serve tiny-random/gpt-oss-mxfp4

• Transformers

import torch
from transformers import pipeline

model_id = "tiny-random/gpt-oss-mxfp4"

pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype='auto',
    device_map="cuda",
)

messages = [
    {"role": "user", "content": "Explain quantum mechanics clearly and concisely."},
]

outputs = pipe(
    messages,
    max_new_tokens=16,
)
print(outputs[0]["generated_text"][-1])

Codes to create this repo:

import json

import safetensors
import torch
from huggingface_hub import hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    AutoTokenizer,
    GenerationConfig,
    GptOssForCausalLM,
    pipeline,
    set_seed,
)

source_model_id = "openai/gpt-oss-120b"
save_folder = "/tmp/tiny-random/gpt-oss-mxfp4"

processor = AutoProcessor.from_pretrained(source_model_id)
processor.save_pretrained(save_folder)

with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model'), 'r') as f:
    config_json = json.load(f)
config_json.update({
    "head_dim": 32,
    "hidden_size": 32,  # required by Mxfp4GptOssExperts codes
    "intermediate_size": 64,
    "layer_types": ["sliding_attention", "full_attention"],
    "num_attention_heads": 2,
    "num_hidden_layers": 2,
    "num_key_value_heads": 1,
    "num_local_experts": 32,
    "tie_word_embeddings": True,
})
quantization_config = config_json['quantization_config']
del config_json['quantization_config']
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(save_folder)
print(config)
torch.set_default_dtype(torch.bfloat16)
model = AutoModelForCausalLM.from_config(config, torch_dtype=torch.bfloat16)
torch.set_default_dtype(torch.float32)
model.generation_config = GenerationConfig.from_pretrained(
    source_model_id, trust_remote_code=True,
)
set_seed(42)
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        print(name, p.shape)
model.save_pretrained(save_folder)

# mxfp4
state_dict = model.cpu().state_dict()
del state_dict['lm_head.weight']
for i in range(len(model.model.layers)):
    del state_dict[f'model.layers.{i}.mlp.experts.down_proj']
    del state_dict[f'model.layers.{i}.mlp.experts.gate_up_proj']
    state_dict[f'model.layers.{i}.mlp.experts.down_proj_blocks'] = torch.randint(0, 255, size=(
        config.num_local_experts, config.hidden_size, config.intermediate_size // 32, 16), dtype=torch.uint8
    )
    state_dict[f'model.layers.{i}.mlp.experts.down_proj_scales'] = torch.randint(0, 4, size=(
        config.num_local_experts, config.hidden_size, config.intermediate_size // 32), dtype=torch.uint8
    )
    state_dict[f'model.layers.{i}.mlp.experts.gate_up_proj_blocks'] = torch.randint(0, 255, size=(
        config.num_local_experts, 2 * config.intermediate_size, config.hidden_size // 32, 16), dtype=torch.uint8
    )
    state_dict[f'model.layers.{i}.mlp.experts.gate_up_proj_scales'] = torch.randint(0, 4, size=(
        config.num_local_experts, 2 * config.intermediate_size, config.hidden_size // 32), dtype=torch.uint8
    )
safetensors.torch.save_file(state_dict, f"{save_folder}/model.safetensors")

# from unittest.mock import Mock
# from transformers.quantizers.auto import AutoHfQuantizer
# from transformers.quantizers.quantizer_mxfp4 import Mxfp4HfQuantizer
# _get_device_capability = torch.cuda.get_device_capability
# torch.cuda.get_device_capability = Mock(return_value=(9, 0))
# set_seed(42)
# bf16_state_dict = model.cpu().state_dict()
# model = AutoModelForCausalLM.from_pretrained(save_folder, torch_dtype=torch.bfloat16, quantization_config=quantization_config)
# for i in range(len(model.model.layers)):
#     model.model.layers[i].mlp.experts.down_proj_bottom_pad = 0
#     model.model.layers[i].mlp.experts.down_proj_right_pad = 0
# hf_quantizer: Mxfp4HfQuantizer = AutoHfQuantizer.from_config(quantization_config)
# hf_quantizer.pre_quantized = False
# ffn_keys = ['model.layers.0.mlp.experts.down_proj', 'model.layers.0.mlp.experts.gate_up_proj',
#             'model.layers.1.mlp.experts.down_proj', 'model.layers.1.mlp.experts.gate_up_proj']
# for key in ffn_keys:
#     hf_quantizer.create_quantized_param(model, bf16_state_dict[key], key, "cuda", bf16_state_dict)
# print('down_proj', model.model.layers[0].mlp.experts.down_proj)
# print('down_proj_blocks', model.model.layers[0].mlp.experts.down_proj_blocks)
# state_dict = model.state_dict()
# del state_dict['lm_head.weight']
# for key in ffn_keys:
#     del state_dict[key]
# for k, v in state_dict.items():
#     if str(v.device) == 'meta':
#         print(k, v.device, v.shape)

# safetensors.torch.save_file(state_dict, f"{save_folder}/model.safetensors")
with open(f"{save_folder}/config.json", "r", encoding='utf-8') as f:
    config = json.load(f)
config['quantization_config'] = quantization_config
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config, f, indent=2)
# torch.cuda.get_device_capability = _get_device_capability