Upload generate-responses.py with huggingface_hub

generate-responses.py

@@ -47,6 +47,7 @@ from torch import cuda
 from tqdm.auto import tqdm
 from transformers import AutoTokenizer
 from vllm import LLM, SamplingParams
+from vllm.sampling_params import GuidedDecodingParams
 
 # Enable HF Transfer for faster downloads
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
@@ -162,7 +163,7 @@ uv run https://huggingface.co/datasets/uv-scripts/vllm/raw/main/generate-respons
 def main(
     src_dataset_hub_id: str,
     output_dataset_hub_id: str,
-    model_id: str = "Qwen/Qwen2.5-72B-Instruct",
+    model_id: str = "Qwen/Qwen3-235B-A22B-Instruct-2507",
     messages_column: str = "messages",
     prompt_column: Optional[str] = None,
     output_column: str = "response",
@@ -233,7 +234,13 @@ def main(
     # Initialize vLLM
     logger.info(f"Loading model: {model_id}")
     if max_model_len is None:
-        max_model_len = 16384  # sensible default; most prompts are <8K tokens
+        max_model_len = 32768  # Qwen2.5-72B default context length
+    # Auto-enable FP8 quantization for large models (>150B) to fit on 4xA100
+    quantization = None
+    if "235B" in model_id or "405B" in model_id:
+        quantization = "fp8"
+        logger.info(f"Auto-enabling FP8 quantization for large model: {model_id}")
+
     vllm_kwargs = {
         "model": model_id,
         "tensor_parallel_size": tensor_parallel_size,
@@ -241,6 +248,8 @@ def main(
         "enable_prefix_caching": True,
         "max_model_len": max_model_len,
     }
+    if quantization:
+        vllm_kwargs["quantization"] = quantization
     logger.info(f"Using max_model_len={max_model_len}")
 
     llm = LLM(**vllm_kwargs)
@@ -250,7 +259,7 @@ def main(
     logger.info("Loading tokenizer...")
     tokenizer = AutoTokenizer.from_pretrained(model_id)
 
-    # Create sampling parameters
+    # Create sampling parameters with guided decoding to enforce valid JSON
     sampling_params = SamplingParams(
         temperature=temperature,
         top_p=top_p,
@@ -258,7 +267,7 @@ def main(
         min_p=min_p,
         max_tokens=max_tokens,
         repetition_penalty=repetition_penalty,
-        # no guided_decoding here
+        guided_decoding=GuidedDecodingParams(json_object=True),
     )
 
     # Load dataset
@@ -444,7 +453,7 @@ Examples:
         parser.add_argument(
             "--model-id",
             type=str,
-            default=os.environ.get("MODEL_ID", "Qwen/Qwen2.5-72B-Instruct"),
+            default=os.environ.get("MODEL_ID", "Qwen/Qwen3-235B-A22B-Instruct-2507"),
             help="Model to use for generation (default: Qwen3-235B-A22B, or MODEL_ID env var)",
         )
         parser.add_argument(