fix(grpo): tail-truncate context inside _to_prompt to satisfy vllm input-length gate

train/grpo.py

@@ -184,13 +184,39 @@ def run_grpo(
         "respond with the shortest exact answer span."
     )
 
+    # vLLM hard-checks final prompt length against the model's max_position_embeddings
+    # (32 768 for Qwen 2.5 Coder 0.5B/1.5B). Our train.jsonl `context` field is the
+    # full long document (often >40K tokens). `max_prompt_length` in GRPOConfig is a
+    # post-tokenization cap that TRL applies AFTER vLLM has already rejected the
+    # request (`vllm.exceptions.VLLMValidationError: prompt contains 37220 tokens`).
+    # We must truncate inside `_to_prompt`, BEFORE chat templating.
+    #
+    # Budget: keep the prompt comfortably under cfg.train.max_prompt_length so the
+    # system message + chat-template overhead don't push us over. Tail-truncate the
+    # context (recent text usually contains the answer span in our synthetic data).
+    max_prompt_tok = int(cfg.train.max_prompt_length)
+    chat_overhead_tok = 256  # system msg + chat template wrappers + question
+    ctx_budget_tok = max(256, max_prompt_tok - chat_overhead_tok)
+
+    def _truncate_to_tokens(text: str, max_tokens: int) -> str:
+        if not text:
+            return ""
+        ids = tokenizer.encode(text, add_special_tokens=False)
+        if len(ids) <= max_tokens:
+            return text
+        # Keep the tail: synthetic gold answers are sampled across the doc, so tail
+        # is no worse than head and is cheaper to slice.
+        return tokenizer.decode(ids[-max_tokens:], skip_special_tokens=True)
+
     def _to_prompt(example: dict) -> dict:
+        ctx_full = example.get("context", "") or ""
+        ctx_truncated = _truncate_to_tokens(ctx_full, ctx_budget_tok)
         msgs = [
             {"role": "system", "content": sys_msg},
             {
                 "role": "user",
                 "content": (
-                    f"Context:\n{example.get('context','')}\n\n"
+                    f"Context:\n{ctx_truncated}\n\n"
                     f"{example.get('prompt','')}"
                 ),
             },