[Perf] Optimize `reshape_and_cache_flash` CUDA Kernel (#22036)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from __future__ import annotations
+
+import random
+import time
+
+import torch
+from tabulate import tabulate
+
+from vllm import _custom_ops as ops
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.utils import (
+    STR_DTYPE_TO_TORCH_DTYPE,
+    FlexibleArgumentParser,
+    create_kv_caches_with_random_flash,
+)
+
+logger = init_logger(__name__)
+
+
+@torch.inference_mode()
+def run_benchmark(
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    num_blocks: int,
+    dtype: torch.dtype,
+    kv_cache_dtype: str,
+    kv_cache_layout: str,
+    num_iters: int,
+    device: str = "cuda",
+) -> float:
+    """Return latency (seconds) for given num_tokens."""
+
+    if kv_cache_dtype == "fp8" and head_size % 16:
+        raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
+
+    current_platform.seed_everything(42)
+    torch.set_default_device(device)
+
+    # create random key / value tensors [T, H, D].
+    key = torch.randn(num_tokens, num_heads, head_size, dtype=dtype, device=device)
+    value = torch.randn_like(key)
+
+    # prepare the slot mapping.
+    # each token is assigned a unique slot in the KV-cache.
+    num_slots = block_size * num_blocks
+    if num_tokens > num_slots:
+        raise ValueError("num_tokens cannot exceed the total number of cache slots")
+    slot_mapping_lst = random.sample(range(num_slots), num_tokens)
+    slot_mapping = torch.tensor(slot_mapping_lst, dtype=torch.long, device=device)
+
+    key_caches, value_caches = create_kv_caches_with_random_flash(
+        num_blocks,
+        block_size,
+        1,  # num_layers
+        num_heads,
+        head_size,
+        kv_cache_dtype,
+        dtype,
+        device=device,
+        cache_layout=kv_cache_layout,
+    )
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # compute per-kernel scaling factors for fp8 conversion (if used).
+    k_scale = (key.amax() / 64.0).to(torch.float32)
+    v_scale = (value.amax() / 64.0).to(torch.float32)
+
+    def run_cuda_benchmark(n_iters: int) -> float:
+        nonlocal key, value, key_cache, value_cache, slot_mapping
+        torch.cuda.synchronize()
+        start = time.perf_counter()
+        for _ in range(n_iters):
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+        torch.cuda.synchronize()
+        end = time.perf_counter()
+        return (end - start) / n_iters
+
+    # warm-up
+    run_cuda_benchmark(3)
+
+    lat = run_cuda_benchmark(num_iters)
+
+    # free tensors to mitigate OOM when sweeping
+    del key, value, key_cache, value_cache, slot_mapping
+    torch.cuda.empty_cache()
+
+    return lat
+
+
+def main(args):
+    rows = []
+    for layout in ["NHD", "HND"]:
+        for exp in range(1, 17):
+            n_tok = 2**exp
+            lat = run_benchmark(
+                num_tokens=n_tok,
+                num_heads=args.num_heads,
+                head_size=args.head_size,
+                block_size=args.block_size,
+                num_blocks=args.num_blocks,
+                dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+                kv_cache_dtype=args.kv_cache_dtype,
+                kv_cache_layout=layout,
+                num_iters=args.iters,
+                device="cuda",
+            )
+            rows.append([n_tok, layout, f"{lat * 1e6:.3f}"])
+
+    print(tabulate(rows, headers=["num_tokens", "layout", "latency (µs)"]))
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+
+    parser.add_argument("--num-heads", type=int, default=128)
+    parser.add_argument(
+        "--head-size",
+        type=int,
+        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+        default=128,
+    )
+    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
+    parser.add_argument("--num-blocks", type=int, default=128 * 512)
+
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["half", "bfloat16", "float"],
+        default="bfloat16",
+    )
+
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=["auto", "fp8"],
+        default="auto",
+    )
+
+    parser.add_argument("--iters", type=int, default=100)
+    args = parser.parse_args()
+
+    main(args)

csrc/cache_kernels.cu

@@ -5,6 +5,7 @@
 #include "cuda_utils.h"
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
+#include "quantization/vectorization_utils.cuh"
 
 #ifdef USE_ROCM
   #include "quantization/fp8/amd/quant_utils.cuh"
@@ -261,14 +262,26 @@ __global__ void reshape_and_cache_kernel(
   }
 }
 
+// Used by vectorization_utils to copy/convert one element
+template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
+struct CopyWithScaleOp {
+  float scale;
+
+  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      dst = static_cast<OutT>(src);
+    } else {
+      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
+    }
+  }
+};
+
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_flash_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
     const scalar_t* __restrict__ value,  // [num_tokens, num_heads, head_size]
-    cache_t* __restrict__ key_cache,     // [num_blocks, block_size, num_heads,
-                                         // head_size]
-    cache_t* __restrict__ value_cache,   // [num_blocks, block_size, num_heads,
-                                         // head_size]
+    cache_t* __restrict__ key_cache,     // NHD or HND, shape see comments below
+    cache_t* __restrict__ value_cache,   // same above
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int64_t block_stride, const int64_t page_stride,
     const int64_t head_stride, const int64_t key_stride,
@@ -282,25 +295,58 @@ __global__ void reshape_and_cache_flash_kernel(
   }
   const int64_t block_idx = slot_idx / block_size;
   const int64_t block_offset = slot_idx % block_size;
-  const int n = num_heads * head_size;
-  for (int i = threadIdx.x; i < n; i += blockDim.x) {
-    const int64_t src_key_idx = token_idx * key_stride + i;
-    const int64_t src_value_idx = token_idx * value_stride + i;
-    const int head_idx = i / head_size;
-    const int head_offset = i % head_size;
-    const int64_t tgt_key_value_idx = block_idx * block_stride +
-                                      block_offset * page_stride +
-                                      head_idx * head_stride + head_offset;
-    scalar_t tgt_key = key[src_key_idx];
-    scalar_t tgt_value = value[src_value_idx];
-    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-      key_cache[tgt_key_value_idx] = tgt_key;
-      value_cache[tgt_key_value_idx] = tgt_value;
-    } else {
-      key_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
-      value_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
+  const int n_elems = num_heads * head_size;
+
+  // pointers to the beginning of the source row for this token.
+  const scalar_t* __restrict__ key_src = key + token_idx * key_stride;
+  const scalar_t* __restrict__ value_src = value + token_idx * value_stride;
+
+  // find the start position inside the kv-cache for this token.
+  cache_t* __restrict__ key_dst =
+      key_cache + block_idx * block_stride + block_offset * page_stride;
+  cache_t* __restrict__ value_dst =
+      value_cache + block_idx * block_stride + block_offset * page_stride;
+
+  // this is true for the NHD layout where `head_stride == head_size`
+  const bool is_contiguous_heads = (head_stride == head_size);
+
+  float k_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *k_scale;
+  float v_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *v_scale;
+  constexpr int VEC_SIZE = (sizeof(scalar_t) == 2) ? 8 : 4;
+  CopyWithScaleOp<cache_t, scalar_t, kv_dt> k_op{k_scale_val};
+  CopyWithScaleOp<cache_t, scalar_t, kv_dt> v_op{v_scale_val};
+  if (is_contiguous_heads) {
+    // NHD layout
+    // kv cache: [num_blocks, block_size, num_heads, head_size]
+    vectorize_with_alignment<VEC_SIZE>(key_src, key_dst, n_elems, threadIdx.x,
+                                       blockDim.x, k_op);
+
+    vectorize_with_alignment<VEC_SIZE>(value_src, value_dst, n_elems,
+                                       threadIdx.x, blockDim.x, v_op);
+
+  } else {
+    // HND layout: heads are strided, but each head_size segment is contiguous
+    // kv cache: [num_blocks, num_heads, block_size, head_size]
+    const int lane = threadIdx.x & 31;     // 0..31 within warp
+    const int warp_id = threadIdx.x >> 5;  // warp index within block
+    const int warps_per_block = blockDim.x >> 5;
+
+    for (int head = warp_id; head < num_heads; head += warps_per_block) {
+      const scalar_t* __restrict__ k_src_h = key_src + head * head_size;
+      const scalar_t* __restrict__ v_src_h = value_src + head * head_size;
+
+      cache_t* __restrict__ k_dst_h =
+          key_dst + static_cast<int64_t>(head) * head_stride;
+      cache_t* __restrict__ v_dst_h =
+          value_dst + static_cast<int64_t>(head) * head_stride;
+
+      // within each head, let the 32 threads of the warp perform the vector
+      // copy
+      vectorize_with_alignment<VEC_SIZE>(k_src_h, k_dst_h, head_size, lane, 32,
+                                         k_op);
+
+      vectorize_with_alignment<VEC_SIZE>(v_src_h, v_dst_h, head_size, lane, 32,
+                                         v_op);
     }
   }
 }