rename minimind→omni in model_type, eval/convert scripts, and docs

README.md

@@ -173,25 +173,37 @@ torchrun --nproc_per_node=4 -m trainers.lm.full_sft --config configs/lm/lm_full_
 ### 推理 / 对话
 
 ```bash
-# 加载 SFT 模型对话（mini 版）
-python scripts/eval_llm.py --save_dir checkpoint/lm_full_sft_mini \
-                           --weight full_sft --hidden_size 128 --native
+# 原生 torch 格式（.pth）
+python scripts/eval_llm.py --native --save_dir checkpoint/lm_full_sft_mini \
+                           --weight full_sft --hidden_size 128
 
-# 加载 pretrain 模型（仅续写，无对话格式）
-python scripts/eval_llm.py --save_dir checkpoint/lm_pretrain_mini \
-                           --weight pretrain --hidden_size 128 --native
+python scripts/eval_llm.py --native --save_dir checkpoint/lm_pretrain_mini \
+                           --weight pretrain --hidden_size 128
 
-# 全量版（hidden_size=512）
-python scripts/eval_llm.py --save_dir checkpoint/lm/full_sft \
-                           --weight full_sft --hidden_size 512 --native
+# HuggingFace 格式（config.json + model.safetensors）
+python scripts/eval_llm.py --load_from checkpoint/omni \
+                           --tokenizer_path checkpoint/omni
 
-# 多模态视觉 VLM
-python scripts/eval_vlm.py --save_dir checkpoint/vlm --weight full_sft
+python scripts/eval_llm.py --load_from checkpoint/lm_full_sft_mini/hf \
+                           --tokenizer_path checkpoint/lm_full_sft_mini/hf
 
-# 全模态 VAM（文本 + 视觉 + 语音）
+# 多模态（VLM / VAM）
+python scripts/eval_vlm.py --save_dir checkpoint/vlm --weight full_sft
 python scripts/eval_vam.py --save_dir checkpoint/vam --weight full_sft
 ```
 
+### 格式转换
+
+```bash
+# 原生 torch → HuggingFace 格式（omni 原生）
+python scripts/convert_model.py checkpoint/lm_full_sft_mini/full_sft_128.pth \
+                               checkpoint/lm_full_sft_mini/hf \
+                               --tokenizer_path checkpoint/tokenizer
+
+# 省略 --mode，默认 omni；也可输出 Qwen3 兼容格式
+python scripts/convert_model.py checkpoint/omni/full_sft.pth output/dir --mode qwen
+```
+
 ## 配置说明
 
 `configs/*.yaml` 分为 `model` / `train` / `paths` 三段，由 `utils.training.apply_config`

scripts/convert_model.py

@@ -10,7 +10,7 @@ from models.lm.lora import apply_lora, merge_lora
 
 warnings.filterwarnings('ignore', category=UserWarning)
 
-def convert_torch2transformers_minimind(torch_path, transformers_path, dtype=torch.float16):
+def convert_torch2transformers_omni(torch_path, transformers_path, dtype=torch.float16, tokenizer_path='checkpoint/tokenizer'):
     LMConfig.register_for_auto_class()
     LMForCausalLM.register_for_auto_class("AutoModelForCausalLM")
     lm_model = LMForCausalLM(lm_config)
@@ -21,7 +21,7 @@ def convert_torch2transformers_minimind(torch_path, transformers_path, dtype=tor
     model_params = sum(p.numel() for p in lm_model.parameters() if p.requires_grad)
     print(f'模型参数: {model_params / 1e6} 百万 = {model_params / 1e9} B (Billion)')
     lm_model.save_pretrained(transformers_path, safe_serialization=False)
-    tokenizer = AutoTokenizer.from_pretrained('../model/')
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
     tokenizer.save_pretrained(transformers_path)
     # ======= transformers-5.0的兼容低版本写法 =======
     if int(transformers.__version__.split('.')[0]) >= 5:
@@ -30,11 +30,11 @@ def convert_torch2transformers_minimind(torch_path, transformers_path, dtype=tor
         config = json.load(open(config_path, 'r', encoding='utf-8'))
         config['rope_theta'] = lm_config.rope_theta; config['rope_scaling'] = None; del config['rope_parameters']
         json.dump(config, open(config_path, 'w', encoding='utf-8'), indent=2, ensure_ascii=False)
-    print(f"模型已保存为 Transformers-MiniMind 格式: {transformers_path}")
+    print(f"模型已保存为 Transformers-Omni 格式: {transformers_path}")
 
 
 # QwenForCausalLM/LlamaForCausalLM结构兼容生态
-def convert_torch2transformers(torch_path, transformers_path, dtype=torch.float16):
+def convert_torch2transformers(torch_path, transformers_path, dtype=torch.float16, tokenizer_path='checkpoint/tokenizer'):
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
     state_dict = torch.load(torch_path, map_location=device)
     common_config = {
@@ -80,7 +80,7 @@ def convert_torch2transformers(torch_path, transformers_path, dtype=torch.float1
     qwen_model.save_pretrained(transformers_path)
     model_params = sum(p.numel() for p in qwen_model.parameters() if p.requires_grad)
     print(f'模型参数: {model_params / 1e6} 百万 = {model_params / 1e9} B (Billion)')
-    tokenizer = AutoTokenizer.from_pretrained('../model/')
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
     tokenizer.save_pretrained(transformers_path)
 
     # ======= transformers-5.0的兼容低版本写法 =======
@@ -123,19 +123,28 @@ def convert_json_to_jinja(json_file_path, output_path):
 
 
 if __name__ == '__main__':
-    lm_config = LMConfig(hidden_size=768, num_hidden_layers=8, max_seq_len=8192, use_moe=False)
-
-    # convert torch to transformers
-    torch_path = f"../checkpoint/full_sft_{lm_config.hidden_size}{'_moe' if lm_config.use_moe else ''}.pth"
-    transformers_path = '../minimind-3'
-    convert_torch2transformers(torch_path, transformers_path)
-
-    # # merge lora
-    # base_torch_path = f"../checkpoint/full_sft_{lm_config.hidden_size}{'_moe' if lm_config.use_moe else ''}.pth"
-    # lora_path = f"../checkpoint/lora_identity_{lm_config.hidden_size}{'_moe' if lm_config.use_moe else ''}.pth"
-    # merged_torch_path = f"../checkpoint/merge_identity_{lm_config.hidden_size}{'_moe' if lm_config.use_moe else ''}.pth"
-    # convert_merge_base_lora(base_torch_path, lora_path, merged_torch_path)
-
-    # convert_transformers2torch(transformers_path, torch_path)
-    # convert_json_to_jinja('../model/tokenizer_config.json', '../model/chat_template.jinja')
-    # convert_jinja_to_json('../model/chat_template.jinja')
+    import argparse
+    parser = argparse.ArgumentParser(description="转换模型格式")
+    parser.add_argument('torch_path', type=str, help="输入 .pth 权重路径")
+    parser.add_argument('output_dir', type=str, help="输出目录")
+    parser.add_argument('--hidden_size', type=int, default=None, help="hidden_size（未指定则从 checkpoint 推断）")
+    parser.add_argument('--num_hidden_layers', type=int, default=None, help="层数（未指定则从 checkpoint 推断）")
+    parser.add_argument('--use_moe', action='store_true', help="MoE 架构")
+    parser.add_argument('--max_seq_len', type=int, default=8192, help="最大序列长度")
+    parser.add_argument('--mode', choices=['omni', 'qwen'], default='omni', help="输出格式（omni=原生LM格式, qwen=Qwen3兼容格式）")
+    parser.add_argument('--dtype', type=str, default='float16', help="权重精度")
+    parser.add_argument('--tokenizer_path', type=str, default='checkpoint/tokenizer', help="tokenizer 路径")
+    args = parser.parse_args()
+
+    state = torch.load(args.torch_path, map_location='cpu')
+    hs = args.hidden_size or state['model.embed_tokens.weight'].shape[1]
+    nl = args.num_hidden_layers or (max(int(k.split('.')[2]) for k in state if k.startswith('model.layers.')) + 1)
+    dtype = getattr(torch, args.dtype)
+
+    lm_config = LMConfig(hidden_size=hs, num_hidden_layers=nl,
+                         max_seq_len=args.max_seq_len, use_moe=args.use_moe)
+
+    if args.mode == 'omni':
+        convert_torch2transformers_omni(args.torch_path, args.output_dir, dtype, args.tokenizer_path)
+    else:
+        convert_torch2transformers(args.torch_path, args.output_dir, dtype, args.tokenizer_path)

scripts/eval_llm.py

@@ -3,7 +3,7 @@ import argparse
 import random
 import warnings
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer
+from transformers import AutoTokenizer, TextStreamer
 from models import LMConfig, LMForCausalLM
 from models.lm.lora import *  # noqa: F401,F403
 from utils.training import setup_seed, get_model_params
@@ -27,7 +27,7 @@ def init_model(args):
             apply_lora(model)
             load_lora(model, f'./{args.save_dir}/{args.lora_weight}_{args.hidden_size}.pth')
     else:
-        model = AutoModelForCausalLM.from_pretrained(args.load_from, trust_remote_code=True)
+        model = LMForCausalLM.from_pretrained(args.load_from)
     get_model_params(model, model.config)
     return model.half().eval().to(args.device), tokenizer
 

scripts/serve_openai_api.py

@@ -218,7 +218,7 @@ async def chat_completions(request: ChatRequest):
                 "id": f"chatcmpl-{int(time.time())}",
                 "object": "chat.completion",
                 "created": int(time.time()),
-                "model": "minimind",
+                "model": "omni",
                 "choices": [
                     {
                         "index": 0,

src/models/lm/config.py

@@ -3,7 +3,7 @@ from transformers import PretrainedConfig
 
 
 class LMConfig(PretrainedConfig):
-    model_type = "minimind"
+    model_type = "omni"
 
     def __init__(self, hidden_size=768, num_hidden_layers=8, use_moe=False, **kwargs):
         super().__init__(**kwargs)
@@ -17,25 +17,35 @@ class LMConfig(PretrainedConfig):
         self.flash_attn = kwargs.get("flash_attn", True)
         self.num_attention_heads = kwargs.get("num_attention_heads", 8)
         self.num_key_value_heads = kwargs.get("num_key_value_heads", 4)
-        self.head_dim = kwargs.get("head_dim", self.hidden_size // self.num_attention_heads)
-        self.hidden_act = kwargs.get("hidden_act", 'silu')
-        self.intermediate_size = kwargs.get("intermediate_size", math.ceil(hidden_size * math.pi / 64) * 64)
+        self.head_dim = kwargs.get(
+            "head_dim", self.hidden_size // self.num_attention_heads
+        )
+        self.hidden_act = kwargs.get("hidden_act", "silu")
+        self.intermediate_size = kwargs.get(
+            "intermediate_size", math.ceil(hidden_size * math.pi / 64) * 64
+        )
         self.max_position_embeddings = kwargs.get("max_position_embeddings", 32768)
         self.rms_norm_eps = kwargs.get("rms_norm_eps", 1e-6)
         self.rope_theta = kwargs.get("rope_theta", 1e6)
         self.tie_word_embeddings = kwargs.get("tie_word_embeddings", True)
         self.inference_rope_scaling = kwargs.get("inference_rope_scaling", False)
-        self.rope_scaling = {
-            "beta_fast": 32,
-            "beta_slow": 1,
-            "factor": 16,
-            "original_max_position_embeddings": 2048,
-            "attention_factor": 1.0,
-            "type": "yarn"
-        } if self.inference_rope_scaling else None
+        self.rope_scaling = (
+            {
+                "beta_fast": 32,
+                "beta_slow": 1,
+                "factor": 16,
+                "original_max_position_embeddings": 2048,
+                "attention_factor": 1.0,
+                "type": "yarn",
+            }
+            if self.inference_rope_scaling
+            else None
+        )
         # MoE specific configs (ignored if use_moe = False)
         self.num_experts = kwargs.get("num_experts", 4)
         self.num_experts_per_tok = kwargs.get("num_experts_per_tok", 1)
-        self.moe_intermediate_size = kwargs.get("moe_intermediate_size", self.intermediate_size)
+        self.moe_intermediate_size = kwargs.get(
+            "moe_intermediate_size", self.intermediate_size
+        )
         self.norm_topk_prob = kwargs.get("norm_topk_prob", True)
         self.router_aux_loss_coef = kwargs.get("router_aux_loss_coef", 5e-4)

src/models/vam/config.py

@@ -2,7 +2,7 @@ from models.lm.config import LMConfig
 
 
 class VAMConfig(LMConfig):
-    model_type = "minimind-o"
+    model_type = "omni-o"
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)

src/models/vlm/config.py

@@ -2,7 +2,7 @@ from models.lm.config import LMConfig
 
 
 class VLMConfig(LMConfig):
-    model_type = "minimind-v"
+    model_type = "omni-v"
 
     def __init__(self, image_special_token='<|image_pad|>', image_ids=[12], **kwargs):
         self.image_special_token = image_special_token

src/trainers/lm/rollout_engine.py

@@ -1,7 +1,7 @@
 """Rollout engines for RL training.
 
 To use the SGLang backend, first launch the server (with a transformers-format model):
-    python -m sglang.launch_server --model-path ./minimind-3 --attention-backend triton --host 0.0.0.0 --port 8998
+    python -m sglang.launch_server --model-path ./checkpoint/omni --attention-backend triton --host 0.0.0.0 --port 8998
 """
 import os
 import requests