Allocate more shared memory to attention kernel (vllm-project#1154)

csrc/attention/attention_kernels.cu

@@ -341,6 +341,9 @@ __global__ void single_query_cached_kv_attention_kernel(
 } // namespace vllm
 
 #define LAUNCH_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS)                        \
+  cudaFuncSetAttribute(                                                                       \
+      vllm::single_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>,   \
+      cudaFuncAttributeMaxDynamicSharedMemorySize, shared_mem_size);                          \
   vllm::single_query_cached_kv_attention_kernel<T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS>        \
   <<<grid, block, shared_mem_size, stream>>>(                                                 \
     out_ptr,                                                                                  \
@@ -401,6 +404,8 @@ void single_query_cached_kv_attention_launcher(
   int padded_max_context_len = ((max_context_len + BLOCK_SIZE - 1) / BLOCK_SIZE) * BLOCK_SIZE;
   int logits_size = padded_max_context_len * sizeof(float);
   int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
+  // Python-side check in vllm.worker.worker._check_if_can_support_max_seq_len
+  // Keep that in sync with the logic here!
   int shared_mem_size = std::max(logits_size, outputs_size);
 
   dim3 grid(num_heads, num_seqs);

csrc/cuda_utils.cpp

@@ -0,0 +1,13 @@
+#include <torch/extension.h>
+
+int get_device_attribute(
+    int attribute,
+    int device_id);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def(
+    "get_device_attribute",
+    &get_device_attribute,
+    "Gets the specified device attribute.");
+}
+

csrc/cuda_utils_kernels.cu

@@ -0,0 +1,14 @@
+int get_device_attribute(
+    int attribute,
+    int device_id)
+{
+    int device, value;
+    if (device_id < 0) {
+        cudaGetDevice(&device);
+    }
+    else {
+        device = device_id;
+    }
+    cudaDeviceGetAttribute(&value, static_cast<cudaDeviceAttr>(attribute), device);
+    return value;
+}

setup.py

@@ -195,6 +195,17 @@ def get_torch_arch_list() -> Set[str]:
 )
 ext_modules.append(quantization_extension)
 
+# Misc. CUDA utils.
+cuda_utils_extension = CUDAExtension(
+    name="vllm.cuda_utils",
+    sources=["csrc/cuda_utils.cpp", "csrc/cuda_utils_kernels.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(cuda_utils_extension)
+
 
 def get_path(*filepath) -> str:
     return os.path.join(ROOT_DIR, *filepath)

tests/kernels/test_attention.py

@@ -7,8 +7,12 @@
 from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
 
 from vllm import attention_ops
+from vllm.utils import get_max_shared_memory_bytes
 
-MAX_SEQ_LEN = 8192
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+# This will change depending on the compute capability.
+# - 512 as a buffer
+MAX_SEQ_LEN = get_max_shared_memory_bytes() // FLOAT32_BYTES - 512
 NUM_BLOCKS = 128  # Arbitrary values for testing
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -135,6 +139,7 @@ def test_single_query_cached_kv_attention(
                                    device="cuda")
 
     context_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    context_lens[-1] = MAX_SEQ_LEN
     max_context_len = max(context_lens)
     context_lens = torch.tensor(context_lens, dtype=torch.int, device="cuda")
 
@@ -243,6 +248,7 @@ def test_multi_query_kv_attention(
     torch.cuda.manual_seed(seed)
 
     seq_lens = random.sample(range(1, MAX_SEQ_LEN), num_seqs)
+    seq_lens[-1] = MAX_SEQ_LEN
     num_tokens = sum(seq_lens)
 
     scale = float(1.0 / (head_size**0.5))

vllm/utils.py

@@ -1,10 +1,12 @@
 import enum
-from platform import uname
 import uuid
+from platform import uname
 
 import psutil
 import torch
 
+from vllm import cuda_utils
+
 
 class Device(enum.Enum):
     GPU = enum.auto()
@@ -25,6 +27,15 @@ def reset(self) -> None:
         self.counter = 0
 
 
+def get_max_shared_memory_bytes(gpu: int = 0) -> int:
+    """Returns the maximum shared memory per thread block in bytes."""
+    # https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
+    cudaDevAttrMaxSharedMemoryPerBlockOptin = 97  # pylint: disable=invalid-name
+    max_shared_mem = cuda_utils.get_device_attribute(
+        cudaDevAttrMaxSharedMemoryPerBlockOptin, gpu)
+    return int(max_shared_mem)
+
+
 def get_gpu_memory(gpu: int = 0) -> int:
     """Returns the total memory of the GPU in bytes."""
     return torch.cuda.get_device_properties(gpu).total_memory

vllm/worker/worker.py

@@ -13,7 +13,7 @@
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import SamplerOutput, SequenceData, SequenceGroupMetadata
 from vllm.worker.cache_engine import CacheEngine
-from vllm.utils import get_gpu_memory
+from vllm.utils import get_gpu_memory, get_max_shared_memory_bytes
 
 
 class Worker:
@@ -136,6 +136,10 @@ def profile_num_available_blocks(
     def init_cache_engine(self, cache_config: CacheConfig) -> None:
         self.cache_config = cache_config
         self.block_size = cache_config.block_size
+
+        _check_if_can_support_max_seq_len(self.scheduler_config.max_model_len,
+                                          self.block_size)
+
         self.cache_engine = CacheEngine(self.cache_config, self.model_config,
                                         self.parallel_config)
         self.cache_events = self.cache_engine.events
@@ -347,3 +351,23 @@ def _pad_to_alignment(x: List[int], multiple_of: int) -> List[int]:
 
 def _pad_to_max(x: List[int], max_len: int) -> List[int]:
     return x + [0] * (max_len - len(x))
+
+
+def _check_if_can_support_max_seq_len(max_seq_len: int,
+                                      block_size: int) -> None:
+    # Follows the logic in
+    # attention_kernels.cu::single_query_cached_kv_attention_launcher
+    max_shared_mem = get_max_shared_memory_bytes()
+    float32_bytes = torch.finfo(torch.float).bits // 8
+    padded_max_seq_len = (
+        (max_seq_len + block_size - 1) / block_size) * block_size
+    # padded_max_seq_len + extra buffer
+    required_shared_mem = (padded_max_seq_len + 512) * float32_bytes
+    if padded_max_seq_len * float32_bytes > max_shared_mem:
+        raise RuntimeError(
+            f"vLLM cannot currently support max_model_len={max_seq_len} "
+            f"with block_size={block_size} on GPU with compute "
+            f"capability {torch.cuda.get_device_capability()} "
+            f"(required shared memory {required_shared_mem} > "
+            f"available shared memory {max_shared_mem}). "
+            "This will be fixed in a future release.")