fix: switch trainer Space to vanilla GRPO path

space/training/requirements.txt

@@ -1,24 +1,14 @@
---extra-index-url https://download.pytorch.org/whl/cu128
-# Strategy: pin to Unsloth 2026.4.8's officially supported A100/Ampere matrix
-# `cu128-ampere-torch2100`. Earlier we tried torch 2.6.0 + cu124, which pulled
-# transformers 5.5 → torchao 0.17 → `torch.utils._pytree.register_constant`,
-# a torch 2.7+ symbol → AttributeError at import. Bumping torch to 2.10.0
-# (highest within Unsloth's `<2.11.0` cap) keeps every transitive happy.
-torch==2.10.0
-torchvision==0.25.0
-torchaudio==2.10.0
-unsloth==2026.4.8
-unsloth_zoo>=2026.4.8
-# Bound transformers to <5.5 to skip a known torchao-integration regression on
-# 5.5.x while staying inside Unsloth's allow-list (>=4.51.3 with the explicit
-# version exclusions Unsloth declares in its requires_dist).
-transformers>=4.51.3,<5.5,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5
-trl>=0.18.2,<=0.24.0,!=0.19.0
-peft>=0.18.0,!=0.11.0
-accelerate>=0.34.1
-datasets>=3.4.1,<4.4.0,!=4.0.*,!=4.1.0
-bitsandbytes>=0.45.5,!=0.46.0,!=0.48.0
-xformers==0.0.34
+--extra-index-url https://download.pytorch.org/whl/cu124
+# Minimal vanilla GRPO stack. Avoid Unsloth, bitsandbytes, xformers, vLLM,
+# torchao, and multi-GPU launch paths for reliability on a single A100 Space.
+torch==2.6.0
+torchvision==0.21.0
+torchaudio==2.6.0
+transformers==4.51.3
+trl==0.18.2
+peft==0.18.0
+accelerate==1.5.0
+datasets==3.4.1
 matplotlib>=3.8.0
 
 # Space control panel + env API deps