[Bugfix][Core] Use torch.cuda.memory_stats() to profile peak memory usage (#9352)

Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>

tests/entrypoints/llm/test_lazy_outlines.py

@@ -26,10 +26,12 @@ def test_lazy_outlines(sample_regex):
     # make sure outlines is not imported
     assert 'outlines' not in sys.modules
 
+    # The second LLM needs to request a higher gpu_memory_utilization because
+    # the first LLM has already allocated a full 30% of the gpu memory.
     llm = LLM(model="facebook/opt-125m",
               enforce_eager=True,
               guided_decoding_backend="lm-format-enforcer",
-              gpu_memory_utilization=0.3)
+              gpu_memory_utilization=0.6)
     sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
     outputs = llm.generate(
         prompts=[

tests/entrypoints/offline_mode/test_offline_mode.py

@@ -44,7 +44,7 @@ def test_offline_mode(llm: LLM, monkeypatch):
         LLM(model=MODEL_NAME,
             max_num_batched_tokens=4096,
             tensor_parallel_size=1,
-            gpu_memory_utilization=0.10,
+            gpu_memory_utilization=0.20,
             enforce_eager=True)
     finally:
         # Reset the environment after the test

tests/worker/test_profile.py

@@ -0,0 +1,69 @@
+import torch
+
+from vllm.engine.arg_utils import EngineArgs
+from vllm.utils import get_distributed_init_method, get_ip, get_open_port
+from vllm.worker.cache_engine import CacheEngine
+from vllm.worker.worker import Worker
+
+
+def test_gpu_memory_profiling():
+    # Tests the gpu profiling that happens in order to determine the number of
+    # KV cache blocks that we can allocate on the GPU.
+    # This test mocks the maximum available gpu memory so that it can run on
+    # any gpu setup.
+
+    # Set up engine args to build a worker.
+    engine_args = EngineArgs(model="facebook/opt-125m",
+                             dtype="half",
+                             load_format="dummy")
+    engine_config = engine_args.create_engine_config()
+    engine_config.cache_config.num_gpu_blocks = 1000
+    engine_config.cache_config.num_cpu_blocks = 1000
+
+    # Create the worker.
+    distributed_init_method = get_distributed_init_method(
+        get_ip(), get_open_port())
+    worker = Worker(
+        model_config=engine_config.model_config,
+        parallel_config=engine_config.parallel_config,
+        scheduler_config=engine_config.scheduler_config,
+        device_config=engine_config.device_config,
+        cache_config=engine_config.cache_config,
+        load_config=engine_config.load_config,
+        local_rank=0,
+        rank=0,
+        distributed_init_method=distributed_init_method,
+        is_driver_worker=True,
+    )
+
+    # Load the model so we can profile it
+    worker.init_device()
+    worker.load_model()
+
+    # Set 10GiB as the total gpu ram to be device-agnostic
+    def mock_mem_info():
+        current_usage = torch.cuda.memory_stats(
+        )["allocated_bytes.all.current"]
+        mock_total_bytes = 10 * 1024**3
+        free = mock_total_bytes - current_usage
+
+        return (free, mock_total_bytes)
+
+    from unittest.mock import patch
+    with patch("torch.cuda.mem_get_info", side_effect=mock_mem_info):
+        gpu_blocks, _ = worker.determine_num_available_blocks()
+
+    # Peak vram usage by torch should be 0.7077 GiB
+    # Non-torch allocations should be 0.0079 GiB
+    # 9.0 GiB should be the utilization target
+    # 8.2843 GiB should be available for the KV cache
+    block_size = CacheEngine.get_cache_block_size(
+        engine_config.cache_config, engine_config.model_config,
+        engine_config.parallel_config)
+
+    expected_blocks = (8.2843 * 1024**3) // block_size
+
+    # Check within a small tolerance for portability
+    # Hardware, kernel, or dependency changes could all affect memory
+    # utilization
+    assert abs(gpu_blocks - expected_blocks) < 5

vllm/worker/worker.py

@@ -217,42 +217,76 @@ class Worker(LocalOrDistributedWorkerBase):
         # 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()
+
+        free_memory_pre_profile, 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()
+        torch.cuda.synchronize()
+
+        self._assert_memory_footprint_increased_during_profiling()
+
+        # 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()
+        # After emptying the torch cache, any other increase in gpu ram should
+        # be from non-torch allocations.
+        non_torch_allocations = free_memory_pre_profile - \
+            torch.cuda.mem_get_info()[0]
+        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.
-        torch.cuda.synchronize()
-        free_gpu_memory, total_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, (
-            "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.")
-
         cache_block_size = self.get_cache_block_size_bytes()
         if cache_block_size == 0:
             num_gpu_blocks = 0
             num_cpu_blocks = 0
         else:
-            num_gpu_blocks = int(
+            num_gpu_blocks = int(available_kv_cache_memory // cache_block_size)
-                (total_gpu_memory * self.cache_config.gpu_memory_utilization -
-                 peak_memory) // cache_block_size)
             num_cpu_blocks = int(self.cache_config.swap_space_bytes //
                                  cache_block_size)
         num_gpu_blocks = max(num_gpu_blocks, 0)
         num_cpu_blocks = max(num_cpu_blocks, 0)
+
+        logger.info(
+            "Memory profiling results: total_gpu_memory=%.2fGiB"
+            " initial_memory_usage=%.2fGiB peak_torch_memory=%.2fGiB"
+            " non_torch_memory=%.2fGiB kv_cache_size=%.2fGiB"
+            " gpu_memory_utilization=%.2f", total_gpu_memory / (1024**3),
+            (total_gpu_memory - free_memory_pre_profile) / (1024**3),
+            (peak_memory - non_torch_allocations) / (1024**3),
+            non_torch_allocations / (1024**3),
+            available_kv_cache_memory / (1024**3),
+            self.cache_config.gpu_memory_utilization)
+
+        # Final cleanup
         if self.model_runner.lora_manager:
             self.model_runner.remove_all_loras()
         gc.collect()
-        torch.cuda.empty_cache()
+
         return num_gpu_blocks, num_cpu_blocks
 
+    def _assert_memory_footprint_increased_during_profiling(self):
+        # NOTE(woosuk): Here we assume that the other processes using the same
+        # GPU did not change their memory usage during the profiling.
+        free_gpu_memory, _ = torch.cuda.mem_get_info()
+        assert self.init_gpu_memory - free_gpu_memory > 0, (
+            "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.")
+
     def initialize_cache(self, num_gpu_blocks: int,
                          num_cpu_blocks: int) -> None:
         """Allocate GPU and CPU KV cache with the specified number of blocks.