[V1][BugFix] Fix Generator construction in greedy + seed case (vllm-p…

…roject#10097)

Signed-off-by: Nick Hill <[email protected]>

vllm/v1/worker/gpu_model_runner.py

@@ -146,7 +146,7 @@ def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
         for req_data in scheduler_output.scheduled_new_reqs:
             req_id = req_data.req_id
             sampling_params = req_data.sampling_params
-            if sampling_params.seed is not None:
+            if sampling_params.sampling_type == SamplingType.RANDOM_SEED:
                 generator = torch.Generator(device=self.device)
                 generator.manual_seed(sampling_params.seed)
             else:
@@ -382,7 +382,8 @@ def execute_model(
                 # Rewind the generator state as if the token was not sampled.
                 generator = self.input_batch.generators.get(i)
                 if generator is not None:
-                    generator.set_offset(generator.get_offset() - 1)
+                    # This relies on cuda-specific torch-internal impl details
+                    generator.set_offset(generator.get_offset() - 4)
 
         if sampler_output.logprob_token_ids is None:
             logprob_token_ids = None