[V1] Fix Compilation config & Enable CUDA graph by default (vllm-proj…

…ect#10528)

Signed-off-by: Woosuk Kwon <[email protected]>

vllm/config.py

@@ -2370,7 +2370,7 @@ def __post_init__(self):
 
         if self.compilation_config is None:
             self.compilation_config = CompilationConfig()
-        if envs.VLLM_USE_V1:
+        if envs.VLLM_USE_V1 and not self.model_config.enforce_eager:
             # NOTE(woosuk): Currently, we use inductor because the piecewise
             # CUDA graphs do not work properly with the custom CUDA kernels.
             # FIXME(woosuk): Disable inductor to reduce the compilation time
@@ -2380,6 +2380,7 @@ def __post_init__(self):
             self.compilation_config.use_inductor = True
             self.compilation_config.pass_config.enable_fusion = False
             self.compilation_config.pass_config.enable_reshape = False
+            self.compilation_config.level = CompilationLevel.PIECEWISE
 
         current_platform.check_and_update_config(self)
 

vllm/v1/worker/gpu_model_runner.py

@@ -1,3 +1,4 @@
+import gc
 import time
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Dict, List, Optional, Set, Tuple
@@ -515,7 +516,25 @@ def load_model(self) -> None:
         logger.info("Loading model weights took %.4f GB",
                     self.model_memory_usage / float(2**30))
 
-    def _dummy_run(self, model: nn.Module, num_tokens: int) -> None:
+    @torch.inference_mode()
+    def _dummy_run(
+        self,
+        model: nn.Module,
+        num_tokens: int,
+        kv_caches: List[torch.Tensor],
+    ) -> torch.Tensor:
+        with set_forward_context(None):
+            hidden_states = model(
+                input_ids=None,
+                positions=self.positions[:num_tokens],
+                kv_caches=kv_caches,
+                attn_metadata=None,
+                inputs_embeds=self.inputs_embeds[:num_tokens])
+        return hidden_states
+
+    def profile_run(self) -> None:
+        # TODO(woosuk): Profile the max memory usage of the encoder and
+        # the encoder cache.
         # use an empty tensor instead of `None`` to force Dynamo to pass
         # it by reference, rather by specializing on the value `None`.
         # the `dtype` argument does not matter, and we use `float32` as
@@ -527,23 +546,17 @@ def _dummy_run(self, model: nn.Module, num_tokens: int) -> None:
             torch.tensor([], dtype=torch.float32, device=self.device)
             for _ in range(self.num_attn_layers)
         ]
-        with set_forward_context(None):  # noqa: SIM117
-            with set_compile_context(self.cudagraph_batch_sizes):
-                # Trigger compilation for general shape.
-                model(input_ids=None,
-                      positions=self.positions,
-                      kv_caches=dummy_kv_caches,
-                      attn_metadata=None,
-                      inputs_embeds=self.inputs_embeds)
-
-    @torch.inference_mode()
-    def profile_run(self) -> None:
-        # TODO(woosuk): Profile the max memory usage of the encoder and
-        # the encoder cache.
-        self._dummy_run(self.model, self.max_num_tokens)
+        with set_compile_context(self.cudagraph_batch_sizes):
+            # Trigger compilation for general shape.
+            hidden_states = self._dummy_run(self.model, self.max_num_tokens,
+                                            dummy_kv_caches)
+        logits = self.model.compute_logits(hidden_states, None)
+        logits = logits[:self.max_num_tokens]
+        # TODO(woosuk): Consider the memory usage of the sampler.
         torch.cuda.synchronize()
+        del hidden_states, logits
+        gc.collect()
 
-    @torch.inference_mode()
     def capture_model(self) -> None:
         if not self.use_cuda_graph:
             logger.warning(
@@ -554,18 +567,11 @@ def capture_model(self) -> None:
         start_time = time.perf_counter()
         start_free_gpu_memory = torch.cuda.mem_get_info()[0]
 
-        with set_forward_context(None):
-            # Trigger CUDA graph capture for specific shapes.
-            # Capture the large shapes first so that the smaller shapes
-            # can reuse the memory pool allocated for the large shapes.
-            for num_tokens in reversed(self.cudagraph_batch_sizes):
-                self.model(
-                    input_ids=None,
-                    positions=self.positions[:num_tokens],
-                    kv_caches=self.kv_caches,
-                    attn_metadata=None,
-                    inputs_embeds=self.inputs_embeds[:num_tokens],
-                )
+        # Trigger CUDA graph capture for specific shapes.
+        # Capture the large shapes first so that the smaller shapes
+        # can reuse the memory pool allocated for the large shapes.
+        for num_tokens in reversed(self.cudagraph_batch_sizes):
+            self._dummy_run(self.model, num_tokens, self.kv_caches)
 
         end_time = time.perf_counter()
         end_free_gpu_memory = torch.cuda.mem_get_info()[0]

vllm/v1/worker/gpu_worker.py

@@ -105,35 +105,48 @@ def determine_num_available_blocks(self) -> Tuple[int, int]:
         # Profile the memory usage of the model and get the maximum number of
         # cache blocks that can be allocated with the remaining free memory.
         torch.cuda.empty_cache()
+        torch.cuda.reset_peak_memory_stats()
 
+        _, total_gpu_memory = torch.cuda.mem_get_info()
         # Execute a forward pass with dummy inputs to profile the memory usage
         # of the model.
         self.model_runner.profile_run()
-
-        # Calculate the number of blocks that can be allocated with the
-        # profiled peak memory.
         torch.cuda.synchronize()
-        free_gpu_memory, total_gpu_memory = torch.cuda.mem_get_info()
+
+        free_gpu_memory, _ = torch.cuda.mem_get_info()
         # NOTE(woosuk): Here we assume that the other processes using the same
         # GPU did not change their memory usage during the profiling.
-        peak_memory = self.init_gpu_memory - free_gpu_memory
-        assert peak_memory > 0, (
+        assert self.init_gpu_memory > free_gpu_memory, (
             "Error in memory profiling. "
             f"Initial free memory {self.init_gpu_memory}, current free memory"
             f" {free_gpu_memory}. This happens when the GPU memory was "
             "not properly cleaned up before initializing the vLLM instance.")
 
+        # Get the peak memory allocation recorded by torch
+        peak_memory = torch.cuda.memory_stats()["allocated_bytes.all.peak"]
+
+        # Check for any memory left around that may have been allocated on the
+        # gpu outside of `torch`. NCCL operations, for example, can use a few
+        # GB during a forward pass
+        torch.cuda.empty_cache()
+        torch_allocated_bytes = torch.cuda.memory_stats(
+        )["allocated_bytes.all.current"]
+        total_allocated_bytes = torch.cuda.mem_get_info(
+        )[1] - torch.cuda.mem_get_info()[0]
+        non_torch_allocations = total_allocated_bytes - torch_allocated_bytes
+        if non_torch_allocations > 0:
+            peak_memory += non_torch_allocations
+        available_kv_cache_memory = (
+            total_gpu_memory * self.cache_config.gpu_memory_utilization -
+            peak_memory)
+
+        # Calculate the number of blocks that can be allocated with the
+        # profiled peak memory.
         cache_block_size = _get_cache_block_size(self.cache_config,
                                                  self.model_config,
                                                  self.parallel_config)
-        num_gpu_blocks = int(
-            (total_gpu_memory * self.cache_config.gpu_memory_utilization -
-             peak_memory) // cache_block_size)
+        num_gpu_blocks = int(available_kv_cache_memory // cache_block_size)
         num_gpu_blocks = max(num_gpu_blocks, 0)
-        # if self.model_runner.lora_manager:
-        #     self.model_runner.remove_all_loras()
-        gc.collect()
-        torch.cuda.empty_cache()
         return num_gpu_blocks, 0
 
     def initialize_cache(self, num_gpu_blocks: int) -> None: