Merge pull request #95 from codelion/fix-entropy-decoding-in-local-server

Fix entropy decoding in local server

Author: codelion

optillm/inference.py

@@ -1257,6 +1257,10 @@ def create(
                     # Handle specialized decoding approaches
                     if decoding:
                         logger.info(f"Using specialized decoding approach: {decoding}")
+
+                        # Ensure model is in eval mode and on correct device
+                        pipeline.current_model.eval()
+                        device = pipeline.current_model.device
 
                         if decoding == "cot_decoding":
                             # Use directly available parameters for CoT
@@ -1284,32 +1288,46 @@ def create(
                             completion_tokens = len(pipeline.tokenizer.encode(result))
 
                         elif decoding == "entropy_decoding":
-                            # Configure generator for entropy decoding
-                            generator = None
-                            if seed is not None:
-                                generator = torch.Generator(device=pipeline.current_model.device)
-                                generator.manual_seed(seed)
-                            
-                            # Use directly available parameters for entropy decoding
 
-                            entropy_params = {
-                                "max_new_tokens": max_tokens if max_tokens is not None else 512,
-                                "temperature": 0.666,
-                                "top_p": 0.90,
-                                "top_k": top_k,
-                                "min_p": min_p,
-                                "generator": generator
-                            }
+                            # Ensure model is using full precision
+                            original_dtype = pipeline.current_model.dtype
+                            pipeline.current_model = pipeline.current_model.to(torch.float32)
+
+                            try:
+                                # Configure generator for entropy decoding
+                                generator = None
+                                if seed is not None:
+                                    generator = torch.Generator(device=device)
+                                    generator.manual_seed(seed)
+                                else:
+                                    generator = torch.Generator(device=device)
+                                    generator.manual_seed(1337)  # Default seed as in original implementation
+                                
+                                # Use directly available parameters for entropy decoding
+                                entropy_params = {
+                                    "max_new_tokens": max_tokens if max_tokens is not None else 4096,
+                                    "temperature": temperature,
+                                    "top_p": top_p,
+                                    "top_k": top_k,
+                                    "min_p": min_p,
+                                    "generator": generator
+                                }
+                                
+                                # Disable autocast and run in full precision
+                                with torch.amp.autocast('cuda', enabled=False), torch.inference_mode():
+                                    result = entropy_decode(
+                                        pipeline.current_model,
+                                        pipeline.tokenizer,
+                                        messages,
+                                        **entropy_params
+                                    )
+                                responses = [result]
+                                logprobs_results = [None]
+                                completion_tokens = len(pipeline.tokenizer.encode(result))
 
-                            result = entropy_decode(
-                                pipeline.current_model,
-                                pipeline.tokenizer,
-                                messages,
-                                **entropy_params
-                            )
-                            responses = [result]
-                            logprobs_results = [None]
-                            completion_tokens = len(pipeline.tokenizer.encode(result))
+                            finally:
+                                # Restore original dtype
+                                pipeline.current_model = pipeline.current_model.to(original_dtype)
 
                         else:
                             raise ValueError(f"Unknown specialized decoding approach: {decoding}")

setup.py

@@ -2,7 +2,7 @@
 
 setup(
     name="optillm",
-    version="0.0.12",
+    version="0.0.13",
     packages=find_packages(),
     py_modules=['optillm'],
     package_data={