Add HuggingFace reference comparison test

tests/compare_hf_reference.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+"""Compare single_shot_inference output against HuggingFace Transformers reference.
+
+This script loads the DeepSeek V4 model using the official HuggingFace
+implementation and processes the same input, comparing intermediate values
+at each step to identify discrepancies in our single_shot_inference.py.
+
+Usage (on B200):
+    source /root/dsv4-nvfp4-workspace/venv/bin/activate
+    cd /root/dsv4-nvfp4-workspace/kernel
+    python tests/compare_hf_reference.py
+"""
+import os, sys, json, math
+import torch
+from pathlib import Path
+
+CHECKPOINT_DIR = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
+
+def main():
+    print("Loading HuggingFace reference model...")
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    
+    # Load with BF16 on CPU (we'll move to GPU as needed)
+    tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT_DIR)
+    
+    # Try loading the model - this may fail if the model class isn't in transformers yet
+    try:
+        model = AutoModelForCausalLM.from_pretrained(
+            CHECKPOINT_DIR,
+            torch_dtype=torch.bfloat16,
+            device_map="auto",
+            trust_remote_code=True,
+        )
+        print("Model loaded successfully!")
+    except Exception as e:
+        print(f"Failed to load model: {e}")
+        print("Trying with trust_remote_code=True and quantization_config bypass...")
+        # The NVFP4 quantization might not be supported by HF yet
+        # Try loading with from_config approach
+        from transformers import AutoConfig
+        config = AutoConfig.from_pretrained(CHECKPOINT_DIR, trust_remote_code=True)
+        print(f"Config loaded: {config.model_type}")
+        print(f"Architectures: {config.architectures}")
+        return
+
+    # Process the same input
+    prompt = "The capital of France is"
+    USER_TOKEN = 128803
+    ASSISTANT_TOKEN = 128804
+    
+    # Build input
+    bos_id = tokenizer.bos_token_id or 0
+    input_ids_list = [bos_id, USER_TOKEN]
+    input_ids_list += tokenizer.encode(prompt, add_special_tokens=False)
+    input_ids_list.append(ASSISTANT_TOKEN)
+    input_ids = torch.tensor([input_ids_list], dtype=torch.long).cuda()
+    
+    print(f"Input: {input_ids.shape[1]} tokens")
+    print(f"Decoded: {tokenizer.decode(input_ids[0][:30])}")
+    
+    # Generate
+    with torch.no_grad():
+        output = model.generate(
+            input_ids,
+            max_new_tokens=10,
+            do_sample=False,
+            temperature=1.0,
+        )
+    
+    generated = output[0, input_ids.shape[1]:]
+    print(f"Generated: {tokenizer.decode(generated)}")
+    
+    # Also get logits for the first position
+    with torch.no_grad():
+        outputs = model(input_ids)
+        logits = outputs.logits
+    
+    # Top-5 at the last position
+    last_logits = logits[0, -1]
+    top5v, top5i = torch.topk(last_logits, 5)
+    print(f"\nTop-5 at last position:")
+    for v, i in zip(top5v, top5i):
+        print(f"  {tokenizer.decode([i.item()])} ({i.item()}, {v.item():.3f})")
+    
+    # Check thinking token
+    think_logit = last_logits[128821].item()
+    print(f"Thinking token (128821) logit: {think_logit:.3f}")
+    print(f"Thinking token rank: {(last_logits > think_logit).sum().item()} / {last_logits.shape[0]}")
+
+
+if __name__ == "__main__":
+    main()