fixed seed bugs

src/numen_scriptorium/inference/qwen.py

@@ -97,10 +97,8 @@ def generate(
         prompt = f"指令：{instruction}\n回答："
 
     inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-    generator = None
     if seed is not None:
-        generator = torch.Generator(device=inputs["input_ids"].device)
-        generator.manual_seed(seed)
+        set_seed(int(seed))
 
     with torch.no_grad():
         outputs = model.generate(
@@ -110,7 +108,6 @@ def generate(
             temperature=temperature,
             top_p=top_p,
             eos_token_id=tokenizer.eos_token_id,
-            generator=generator,
         )
     text = tokenizer.decode(outputs[0], skip_special_tokens=True)
     if "回答：" in text:
@@ -136,10 +133,6 @@ def stream_generate(
         prompt = f"指令：{instruction}\n回答："
 
     inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-    generator = None
-    if seed is not None:
-        generator = torch.Generator(device=inputs["input_ids"].device)
-        generator.manual_seed(seed)
 
     streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
     class _EventStoppingCriteria(StoppingCriteria):
@@ -158,18 +151,30 @@ def stream_generate(
         eos_token_id=tokenizer.eos_token_id,
         streamer=streamer,
     )
-    if generator is not None:
-        generate_kwargs["generator"] = generator
     if stop_event is not None:
         generate_kwargs["stopping_criteria"] = StoppingCriteriaList([_EventStoppingCriteria(stop_event)])
 
-    worker = Thread(target=model.generate, kwargs=generate_kwargs)
+    worker_error: list[Exception] = []
+
+    def _run_generate():
+        try:
+            if seed is not None:
+                set_seed(int(seed))
+            model.generate(**generate_kwargs)
+        except Exception as exc:
+            worker_error.append(exc)
+            # Ensure streamer consumer can exit even if generation fails early.
+            streamer.end()
+
+    worker = Thread(target=_run_generate)
     worker.start()
     for new_text in streamer:
         if stop_event is not None and stop_event.is_set():
             break
         yield new_text
     worker.join(timeout=0.5)
+    if worker_error:
+        raise worker_error[0]
 
 
 def get_model_device(model) -> str: