Add batched RoPE kernel (#3095)

tterrysun · web-flow · commit 7e9bd08f60a4 · 2024-03-13T13:45:26.000-07:00
diff --git a/benchmarks/kernels/benchmark_rope.py b/benchmarks/kernels/benchmark_rope.py
@@ -0,0 +1,120 @@
+from typing import Optional
+
+import argparse
+import torch
+import nvtx
+from itertools import accumulate
+from vllm.model_executor.layers.rotary_embedding import get_rope
+
+
+def benchmark_rope_kernels_multi_lora(
+    is_neox_style: bool,
+    batch_size: int,
+    seq_len: int,
+    num_heads: int,
+    head_size: int,
+    rotary_dim: Optional[int],
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+    max_position: int = 8192,
+    base: int = 10000,
+) -> None:
+    torch.random.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+    torch.set_default_device(device)
+    if rotary_dim is None:
+        rotary_dim = head_size
+    # silulating serving 4 LoRAs
+    scaling_factors = [1, 2, 4, 8]
+    # batched RoPE can take multiple scaling factors
+    batched_rope = get_rope(head_size, rotary_dim, max_position, base,
+                            is_neox_style, {
+                                "type": "linear",
+                                "factor": tuple(scaling_factors)
+                            })
+    # non-batched RoPE takes only one scaling factor, we create multiple
+    # instances to simulate the same behavior
+    non_batched_ropes = []
+    for scaling_factor in scaling_factors:
+        non_batched_ropes.append(
+            get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
+                     {
+                         "type": "linear",
+                         "factor": (scaling_factor, )
+                     }))
+
+    positions = torch.randint(0, max_position, (batch_size, seq_len))
+    query = torch.randn(batch_size,
+                        seq_len,
+                        num_heads * head_size,
+                        dtype=dtype)
+    key = torch.randn_like(query)
+
+    # create query offsets for batched RoPE, we concat multiple kv cache
+    # together and each query needs to find the right kv cache of its type
+    offset_map = torch.tensor(
+        list(
+            accumulate([0] + [
+                max_position * scaling_factor * 2
+                for scaling_factor in scaling_factors[:-1]
+            ])))
+    query_types = torch.randint(0,
+                                len(scaling_factors), (batch_size, seq_len),
+                                device=device)
+    # map query types to offsets
+    query_offsets = offset_map[query_types]
+    # the kernel takes flattened offsets
+    flatten_offsets = query_offsets.flatten()
+
+    # batched queries of the same type together for non-batched RoPE
+    queries = [query[query_types == i] for i in range(len(scaling_factors))]
+    keys = [key[query_types == i] for i in range(len(scaling_factors))]
+    packed_qkr = zip(queries, keys, non_batched_ropes)
+    # synchronize before start timing
+    torch.cuda.synchronize()
+    with nvtx.annotate("non-batched", color="yellow"):
+        for q, k, r in packed_qkr:
+            r.forward(positions, q, k)
+    torch.cuda.synchronize()
+    with nvtx.annotate("batched", color="green"):
+        batched_rope.forward(positions, query, key, flatten_offsets)
+    torch.cuda.synchronize()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description="Benchmark the rotary embedding kernels.")
+    parser.add_argument("--is-neox-style", type=bool, default=True)
+    parser.add_argument("--batch-size", type=int, default=16)
+    parser.add_argument("--seq-len", type=int, default=512)
+    parser.add_argument("--num-heads", type=int, default=8)
+    parser.add_argument("--head-size",
+                        type=int,
+                        choices=[64, 80, 96, 112, 128, 256],
+                        default=128)
+    parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32)
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["bfloat16", "float"],
+                        default="float")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--device",
+                        type=str,
+                        choices=["cuda:0", "cuda:1"],
+                        default="cuda:0")
+    args = parser.parse_args()
+    print(args)
+
+    benchmark_rope_kernels_multi_lora(
+        is_neox_style=args.is_neox_style,
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        num_heads=args.num_heads,
+        head_size=args.head_size,
+        rotary_dim=args.rotary_dim,
+        dtype=getattr(torch, args.dtype),
+        seed=args.seed,
+        device=args.device,
+    )
diff --git a/csrc/ops.h b/csrc/ops.h
@@ -53,6 +53,16 @@ void rotary_embedding(
   torch::Tensor& cos_sin_cache,
   bool is_neox);
 
+void batched_rotary_embedding(
+  torch::Tensor& positions,
+  torch::Tensor& query,
+  torch::Tensor& key,
+  int head_size,
+  torch::Tensor& cos_sin_cache,
+  bool is_neox,
+  int rot_dim,
+  torch::Tensor& cos_sin_cache_offsets);
+
 void silu_and_mul(
   torch::Tensor& out,
   torch::Tensor& input);
diff --git a/csrc/pos_encoding_kernels.cu b/csrc/pos_encoding_kernels.cu
@@ -8,7 +8,7 @@
 namespace vllm {
 
 template<typename scalar_t, bool IS_NEOX>
-inline __device__ void apply_rotary_embedding(
+inline __device__ void apply_token_rotary_embedding(
   scalar_t* __restrict__ arr,
   const scalar_t* __restrict__ cos_ptr,
   const scalar_t* __restrict__ sin_ptr,
@@ -38,22 +38,18 @@ inline __device__ void apply_rotary_embedding(
 }
 
 template<typename scalar_t, bool IS_NEOX>
-__global__ void rotary_embedding_kernel(
-  const int64_t* __restrict__ positions,        // [batch_size, seq_len] or [num_tokens]
+inline __device__ void apply_rotary_embedding(
   scalar_t* __restrict__ query,                 // [batch_size, seq_len, num_heads, head_size] or [num_tokens, num_heads, head_size]
   scalar_t* __restrict__ key,                   // [batch_size, seq_len, num_kv_heads, head_size] or [num_tokens, num_kv_heads, head_size]
-  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, rot_dim // 2]
-  const int rot_dim,
-  const int64_t query_stride,
-  const int64_t key_stride,
+  const scalar_t* cache_ptr,
+  const int head_size,
   const int num_heads,
   const int num_kv_heads,
-  const int head_size) {
-  // Each thread block is responsible for one token.
-  const int token_idx = blockIdx.x;
-  int64_t pos = positions[token_idx];
-  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
-
+  const int rot_dim,
+  const int token_idx,
+  const int64_t query_stride,
+  const int64_t key_stride)
+{
   const int embed_dim = rot_dim / 2;
   const scalar_t* cos_ptr = cache_ptr;
   const scalar_t* sin_ptr = cache_ptr + embed_dim;
@@ -63,7 +59,7 @@ __global__ void rotary_embedding_kernel(
     const int head_idx = i / embed_dim;
     const int64_t token_head = token_idx * query_stride + head_idx * head_size;
     const int rot_offset = i % embed_dim;
-    apply_rotary_embedding<scalar_t, IS_NEOX>(query + token_head, cos_ptr,
+    apply_token_rotary_embedding<scalar_t, IS_NEOX>(query + token_head, cos_ptr,
                                               sin_ptr, rot_offset, embed_dim);
   }
 
@@ -72,11 +68,53 @@ __global__ void rotary_embedding_kernel(
     const int head_idx = i / embed_dim;
     const int64_t token_head = token_idx * key_stride + head_idx * head_size;
     const int rot_offset = i % embed_dim;
-    apply_rotary_embedding<scalar_t, IS_NEOX>(key + token_head, cos_ptr,
+    apply_token_rotary_embedding<scalar_t, IS_NEOX>(key + token_head, cos_ptr,
                                               sin_ptr, rot_offset, embed_dim);
   }
 }
 
+template<typename scalar_t, bool IS_NEOX>
+__global__ void rotary_embedding_kernel(
+  const int64_t* __restrict__ positions,        // [batch_size, seq_len] or [num_tokens]
+  scalar_t* __restrict__ query,                 // [batch_size, seq_len, num_heads, head_size] or [num_tokens, num_heads, head_size]
+  scalar_t* __restrict__ key,                   // [batch_size, seq_len, num_kv_heads, head_size] or [num_tokens, num_kv_heads, head_size]
+  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, rot_dim // 2]
+  const int rot_dim,
+  const int64_t query_stride,
+  const int64_t key_stride,
+  const int num_heads,
+  const int num_kv_heads,
+  const int head_size) {
+  // Each thread block is responsible for one token.
+  const int token_idx = blockIdx.x;
+  int64_t pos = positions[token_idx];
+  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
+
+  apply_rotary_embedding<scalar_t, IS_NEOX>(query, key, cache_ptr, head_size, num_heads, num_kv_heads, rot_dim, token_idx, query_stride, key_stride);
+}
+
+template<typename scalar_t, bool IS_NEOX>
+__global__ void batched_rotary_embedding_kernel(
+  const int64_t* __restrict__ positions,              // [batch_size, seq_len] or [num_tokens]
+  scalar_t* __restrict__ query,                       // [batch_size, seq_len, num_heads, head_size] or [num_tokens, num_heads, head_size]
+  scalar_t* __restrict__ key,                         // [batch_size, seq_len, num_kv_heads, head_size] or [num_tokens, num_kv_heads, head_size]
+  const scalar_t* __restrict__ cos_sin_cache,         // [max_position, 2, rot_dim // 2]
+  const int64_t* __restrict__ cos_sin_cache_offsets,  // [batch_size, seq_len] or [num_tokens]
+  const int rot_dim,
+  const int64_t query_stride,
+  const int64_t key_stride,
+  const int num_heads,
+  const int num_kv_heads,
+  const int head_size) {
+  // Each thread block is responsible for one token.
+  const int token_idx = blockIdx.x;
+  int64_t pos = positions[token_idx];
+  int64_t cos_sin_cache_offset = cos_sin_cache_offsets[token_idx];
+  const scalar_t* cache_ptr = cos_sin_cache + (cos_sin_cache_offset + pos) * rot_dim;
+
+  apply_rotary_embedding<scalar_t, IS_NEOX>(query, key, cache_ptr, head_size, num_heads, num_kv_heads, rot_dim, token_idx, query_stride, key_stride);
+}
+
 } // namespace vllm
 
 void rotary_embedding(
@@ -128,3 +166,61 @@ void rotary_embedding(
       }
     });
 }
+
+/*
+Batched version of rotary embedding, pack multiple LoRAs together
+and process in batched manner.
+*/
+void batched_rotary_embedding(
+  torch::Tensor& positions,         // [batch_size, seq_len] or [num_tokens]
+  torch::Tensor& query,             // [batch_size, seq_len, num_heads * head_size] or [num_tokens, num_heads * head_size]
+  torch::Tensor& key,               // [batch_size, seq_len, num_kv_heads * head_size] or [num_tokens, num_kv_heads * head_size]
+  int head_size,
+  torch::Tensor& cos_sin_cache,     // [max_position, rot_dim]
+  bool is_neox,
+  int rot_dim,
+  torch::Tensor& cos_sin_cache_offsets // [num_tokens]
+) {
+  int64_t num_tokens = cos_sin_cache_offsets.size(0);
+  int num_heads = query.size(-1) / head_size;
+  int num_kv_heads = key.size(-1) / head_size;
+  int64_t query_stride = query.stride(-2);
+  int64_t key_stride = key.stride(-2);
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(num_heads * rot_dim / 2, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(
+    query.scalar_type(),
+    "rotary_embedding",
+    [&] {
+      if (is_neox) {
+        vllm::batched_rotary_embedding_kernel<scalar_t, true><<<grid, block, 0, stream>>>(
+          positions.data_ptr<int64_t>(),
+          query.data_ptr<scalar_t>(),
+          key.data_ptr<scalar_t>(),
+          cos_sin_cache.data_ptr<scalar_t>(),
+          cos_sin_cache_offsets.data_ptr<int64_t>(),
+          rot_dim,
+          query_stride,
+          key_stride,
+          num_heads,
+          num_kv_heads,
+          head_size);
+      } else {
+        vllm::batched_rotary_embedding_kernel<scalar_t, false><<<grid, block, 0, stream>>>(
+          positions.data_ptr<int64_t>(),
+          query.data_ptr<scalar_t>(),
+          key.data_ptr<scalar_t>(),
+          cos_sin_cache.data_ptr<scalar_t>(),
+          cos_sin_cache_offsets.data_ptr<int64_t>(),
+          rot_dim,
+          query_stride,
+          key_stride,
+          num_heads,
+          num_kv_heads,
+          head_size);
+      }
+    });
+}
diff --git a/csrc/pybind.cpp b/csrc/pybind.cpp
@@ -56,6 +56,11 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     &rotary_embedding,
     "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
 
+  ops.def(
+    "batched_rotary_embedding",
+    &batched_rotary_embedding,
+    "Apply GPT-NeoX or GPT-J style rotary embedding to query and key (supports multiple loras)");
+
 // Quantization ops
 #ifndef USE_ROCM
   ops.def("awq_gemm", &awq_gemm, "Quantized GEMM for AWQ");
diff --git a/tests/kernels/test_pos_encoding.py b/tests/kernels/test_pos_encoding.py
diff --git a/vllm/model_executor/layers/rotary_embedding.py b/vllm/model_executor/layers/rotary_embedding.py
