.

README.txt

@@ -1,8 +1,9 @@
-mamba env create -f hf-train-cu118.yaml --strict-channel-priority
+mamba env create -f env.yaml --strict-channel-priority
 mamba activate hf-train-cu118
+python -m pip install --no-deps -r requirements-hf.txt
 
-pip install --upgrade pip
+# deepspeed（选一）
-pip install --no-deps -r requirements-hf.txt --constraint constraints-cu118.txt
+DS_BUILD_OPS=0 python -m pip install "deepspeed==0.14.*"
-
+# 或
-# 需要 deepspeed 时再装：
+mamba install -y cuda-toolkit=11.8 cmake ninja
-DS_BUILD_OPS=0 pip install "deepspeed==0.14.*"  # 先不编译 CUDA 内核
+python -m pip install "deepspeed==0.14.*"

constraints-cu118.txt

@@ -1,3 +0,0 @@
-torch==2.1.2
-torchvision==0.16.2
-torchaudio==2.1.2

hf-train-cu118.yaml

@@ -1,32 +1,31 @@
 name: hf-train-cu118
-channels:
+channels: [pytorch, nvidia, conda-forge]
-  - pytorch
-  - nvidia
-  - conda-forge
 dependencies:
   - python=3.10
   - pip
-  # ---- Torch 栈：固定 2.1.2 + cu118 ----
+  # ---- Torch 栈：2.1.2 + cu118 ----
   - pytorch=2.1.2
   - torchvision=0.16.2
   - torchaudio=2.1.2
   - pytorch-cuda=11.8
-  # ---- 避坑：Numpy 钉在 1.26.* ----
+
+  # ---- 数值栈 ----
   - numpy=1.26.*
-  # ---- 常用科学/系统库 ----
   - pandas
   - scipy
   - pyarrow
-  - uvicorn
+
-  - git
+  # ---- HF 主栈 ----
-  # ---- HF 主栈 + 其运行时依赖（全部走 conda，不让 pip 动依赖）----
+  - transformers>=4.40,<5
-  - transformers>=4.40
   - accelerate>=0.30
   - datasets>=2.18
   - evaluate>=0.4
   - safetensors>=0.4
   - sentencepiece>=0.1.99
   - tokenizers=0.19.*
+
+  # ---- 依赖/工具 ----
+  - protobuf<5
   - huggingface_hub>=0.23
   - tqdm>=4.66
   - scikit-learn>=1.4
@@ -41,4 +40,9 @@ dependencies:
   - xxhash
   - aiohttp
   - psutil
-
+  # 可选（按需）
+  # - einops
+  # - ninja
+  # - cmake
+  # - xformers==0.0.22.post3  # 若真需要
+  # - fastapi                # 若要配合 uvicorn 跑服务

requirements-hf.txt

@@ -1,15 +1,5 @@
-# requirements-hf.txt  （HF 生态）
-transformers>=4.40,<5.0
-accelerate>=0.30
-datasets>=2.18
-evaluate>=0.4
-safetensors>=0.4
-sentencepiece>=0.1.99
-tokenizers>=0.19,<0.21
-huggingface_hub>=0.23
-tqdm>=4.66
 peft>=0.11
 bitsandbytes>=0.43
-tensorboard>=2.16
+# 可选：
-scikit-learn>=1.4
+# xformers==0.0.22.post3
-# deepspeed 单独装，别放进来
+# flash-attn