[V1][Core] Fix memory issue with logits & sampling (#13776)

Signed-off-by: Roger Wang <ywang@roblox.com>

tests/basic_correctness/test_cumem.py

@@ -142,7 +142,16 @@ def test_end_to_end(model: str, use_v1: bool):
     used_bytes = total - free_gpu_bytes_after_sleep - used_bytes_baseline
     # now the memory usage is mostly cudagraph memory pool,
     # and it should be less than the model weights (1B model, 2GiB weights)
-    assert used_bytes < 2 * GiB_bytes
+
+    # NOTE: In V1, the memory buffer for logits (max_num_reqs x vocab_size)
+    # is captured but cannot be releasesd from PyTorch due to a known bug,
+    # therefore high memory usage after `llm.sleep` is called is expected.
+    # FIXME(youkaichao & ywang96): Fix memory buffer issue with sleep mode
+    # in V1.
+    if use_v1:
+        assert used_bytes < 7 * GiB_bytes
+    else:
+        assert used_bytes < 2 * GiB_bytes
 
     llm.wake_up()
     output2 = llm.generate(prompt, sampling_params)

vllm/v1/worker/gpu_model_runner.py

@@ -1238,6 +1238,42 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             )
         return hidden_states
 
+    @torch.inference_mode()
+    def _dummy_sampler_run(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+
+        logits = self.model.compute_logits(hidden_states, None)
+        num_reqs = logits.size(0)
+
+        dummy_tensors = lambda v: torch.full(
+            (num_reqs, ), v, device=self.device)
+
+        dummy_metadata = SamplingMetadata(
+            temperature=dummy_tensors(0.5),
+            all_greedy=False,
+            all_random=False,
+            top_p=dummy_tensors(0.9),
+            top_k=dummy_tensors(logits.size(1) - 1),
+            min_p=None,
+            generators={},
+            max_num_logprobs=None,
+            no_penalties=True,
+            prompt_token_ids=None,
+            frequency_penalties=dummy_tensors(0.1),
+            presence_penalties=dummy_tensors(0.1),
+            repetition_penalties=dummy_tensors(0.1),
+            output_token_ids=[[] for _ in range(num_reqs)],
+            min_tokens={},
+            logit_bias=[None for _ in range(num_reqs)],
+            allowed_token_ids_mask=None,
+        )
+        sampler_output = self.model.sample(logits=logits,
+                                           sampling_metadata=dummy_metadata)
+
+        return sampler_output
+
     def profile_run(self) -> None:
         # Profile with multimodal encoder & encoder cache.
         # TODO: handle encoder-decoder models once we support them.
@@ -1353,37 +1389,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             hidden_states = self._dummy_run(self.max_num_tokens)
             if get_pp_group().is_last_rank:
                 hidden_states = hidden_states[logit_indices]
-                logits = self.model.compute_logits(hidden_states, None)
-                dummy_tensors = lambda v: torch.full(
-                    (num_reqs, ), v, device=self.device)
-                dummy_metadata = SamplingMetadata(
-                    temperature=dummy_tensors(0.5),
-                    all_greedy=False,
-                    all_random=False,
-                    top_p=dummy_tensors(0.9),
-                    top_k=dummy_tensors(logits.size(1) - 1),
-                    min_p=None,
-                    generators={},
-                    max_num_logprobs=None,
-                    no_penalties=True,
-                    prompt_token_ids=torch.ones_like(logits,
-                                                     dtype=torch.int64),
-                    frequency_penalties=dummy_tensors(0.1),
-                    presence_penalties=dummy_tensors(0.1),
-                    repetition_penalties=dummy_tensors(0.1),
-                    output_token_ids=[[] for _ in range(num_reqs)],
-                    min_tokens={},
-                    logit_bias=[None for _ in range(num_reqs)],
-                    allowed_token_ids_mask=None,
-                )
-                sampler_output = self.model.sample(
-                    logits=logits, sampling_metadata=dummy_metadata)
+                sampler_output = self._dummy_sampler_run(hidden_states)
             else:
-                logits = None
                 sampler_output = None
-                dummy_metadata = None
             torch.cuda.synchronize()
-            del hidden_states, logits, sampler_output, dummy_metadata
+            del hidden_states, sampler_output
             self.encoder_cache.clear()
         gc.collect()
 

vllm/v1/worker/gpu_worker.py

@@ -119,6 +119,8 @@ class Worker(WorkerBase):
         self.model_runner: GPUModelRunner = GPUModelRunner(
             self.vllm_config, self.device)
 
+    # FIXME(youkaichao & ywang96): Use TorchDispatchMode instead of memory pool
+    # to hijack tensor allocation.
     def load_model(self) -> None:
         if self.vllm_config.model_config.enable_sleep_mode:
             allocator = CuMemAllocator.get_instance()
@@ -211,6 +213,25 @@ class Worker(WorkerBase):
             self.model_runner._dummy_run(size)
         if not self.model_config.enforce_eager:
             self.model_runner.capture_model()
+
+        # Warm up sampler and preallocate memory buffer for logits and other
+        # sampling related tensors of max possible shape to avoid memory
+        # fragmentation issue.
+        # NOTE: This is called after `capture_model` on purpose to prevent
+        # memory buffers from being cleared by `torch.cuda.empty_cache`.
+        try:
+            self.model_runner._dummy_sampler_run(
+                hidden_states=self.model_runner._dummy_run(
+                    num_tokens=self.scheduler_config.max_num_seqs))
+        except RuntimeError as e:
+            if 'out of memory' in str(e):
+                raise RuntimeError(
+                    "CUDA out of memory occurred when warming up sampler. "
+                    "Please try lowering `gpu_memory_utilization` when "
+                    "initializing the engine.") from None
+            else:
+                raise e
+
         # Reset the seed to ensure that the random state is not affected by
         # the model initialization and profiling.
         set_random_seed(self.model_config.seed)