deploy: update train/requirements.txt

train/requirements.txt

@@ -1,49 +1,49 @@
-# Training deps for DebateFloor GRPO on HF Jobs.
-# Tested image: pytorch/pytorch:2.4.0-cuda12.1-cudnn9-runtime
-# (PyTorch 2.4.0, torchvision 0.19.0, CUDA 12.1, Python 3.11)
-
-# Core RL trainer.
-# - GRPO was added in trl 0.13 (Jan 2025) and the GRPOConfig API used in
-#   train/train_minimal.py (processing_class, reward_funcs, num_generations,
-#   max_completion_length, max_prompt_length) stabilized in 0.15.
-# - Cap below 0.20 to keep transformers requirement at 4.48 (avoids needing
-#   torch>=2.5 which the base image doesn't have).
-trl>=0.15.0,<0.18.0
-
-# transformers must be < 4.48 — that's the version where loss_deformable_detr.py
-# was added, which unconditionally imports image_transforms -> torchvision at
-# module load time. Even purging torchvision can't avoid this if transformers
-# tries to import it during _LazyModule resolution.
-# trl 0.15.2 accepts transformers >= 4.46, so 4.46 / 4.47 are both valid.
-transformers>=4.46.0,<4.48.0
-
-# Note: torchvision is INTENTIONALLY not pinned. The HF Jobs base image ships
-# torch 2.11.0+cu130 (not 2.4.0 as the tag suggests), and ABI-matching
-# torchvision wheels for cu130 are not always available. Since DebateFloor is
-# text-only and never uses image transforms, jobs_run.py uninstalls torchvision
-# entirely after pip install — transformers will then skip the broken
-# `image_utils → torchvision::nms` import path.
-
-# Model + LoRA.
-peft>=0.13.0
-accelerate>=1.0.0,<2.0.0
-datasets>=2.19.0
-
-# bitsandbytes prebuilt wheel for CUDA 12.x.
-bitsandbytes>=0.43.0
-
-# Logging / IO
-wandb>=0.17.0
-requests>=2.32.0
-matplotlib>=3.8.0
-
-# Env server (uvicorn + FastAPI come from root requirements.txt; we just need
-# the openenv-core import path because train_minimal.py loads
-# server.calibration_grader + server.claim_generator).
-uvicorn>=0.30.0
-openenv-core>=0.2.3
-
-# Unsloth 4-bit QLoRA (MR-3 hackathon constraint).
-# unsloth-zoo is the runtime backend; pinning a recent stable version known to
-# work with torch 2.4.
-unsloth>=2024.10
+# Training deps for DebateFloor GRPO on HF Jobs.
+# Tested image: pytorch/pytorch:2.4.0-cuda12.1-cudnn9-runtime
+# (PyTorch 2.4.0, torchvision 0.19.0, CUDA 12.1, Python 3.11)
+
+# Core RL trainer.
+# - GRPO was added in trl 0.13 (Jan 2025) and the GRPOConfig API used in
+#   train/train_minimal.py (processing_class, reward_funcs, num_generations,
+#   max_completion_length, max_prompt_length) stabilized in 0.15.
+# - Cap below 0.20 to keep transformers requirement at 4.48 (avoids needing
+#   torch>=2.5 which the base image doesn't have).
+trl>=0.15.0,<0.18.0
+
+# transformers must be < 4.48 — that's the version where loss_deformable_detr.py
+# was added, which unconditionally imports image_transforms -> torchvision at
+# module load time. Even purging torchvision can't avoid this if transformers
+# tries to import it during _LazyModule resolution.
+# trl 0.15.2 accepts transformers >= 4.46, so 4.46 / 4.47 are both valid.
+transformers>=4.46.0,<4.48.0
+
+# Note: torchvision is INTENTIONALLY not pinned. The HF Jobs base image ships
+# torch 2.11.0+cu130 (not 2.4.0 as the tag suggests), and ABI-matching
+# torchvision wheels for cu130 are not always available. Since DebateFloor is
+# text-only and never uses image transforms, jobs_run.py uninstalls torchvision
+# entirely after pip install — transformers will then skip the broken
+# `image_utils → torchvision::nms` import path.
+
+# Model + LoRA.
+peft>=0.13.0
+accelerate>=1.0.0,<2.0.0
+datasets>=2.19.0
+
+# bitsandbytes prebuilt wheel for CUDA 12.x.
+bitsandbytes>=0.43.0
+
+# Logging / IO
+wandb>=0.17.0
+requests>=2.32.0
+matplotlib>=3.8.0
+
+# Env server (uvicorn + FastAPI come from root requirements.txt; we just need
+# the openenv-core import path because train_minimal.py loads
+# server.calibration_grader + server.claim_generator).
+uvicorn>=0.30.0
+openenv-core>=0.2.3
+
+# Unsloth 4-bit QLoRA (MR-3 hackathon constraint).
+# unsloth-zoo is the runtime backend; pinning a recent stable version known to
+# work with torch 2.4.
+unsloth>=2024.10