Implement custom kernel for LLaMA rotary embedding

cacheflow/models/attention.py

@@ -6,13 +6,14 @@
 
 from cacheflow import attention_ops
 from cacheflow import cache_ops
+from cacheflow import pos_encoding_ops
 from cacheflow.models import InputMetadata
 
 
-class OPTCacheFlowAttention(nn.Module):
+class GPTCacheFlowAttention(nn.Module):
 
     def __init__(self, scale: float) -> None:
-        super(OPTCacheFlowAttention, self).__init__()
+        super().__init__()
         self.scale = float(scale)
 
         self.flash_attn = FlashAttention(softmax_scale=self.scale)
@@ -136,3 +137,71 @@ def forward(
         # Reshape the output tensor.
         # NOTE(woosuk): The output tensor may include paddings.
         return output.view(-1, num_heads * head_size)
+
+
+class OPTCacheFlowAttention(GPTCacheFlowAttention):
+    """OPT uses the same attention mechanism as GPT."""
+
+    def __init__(self, scale: float) -> None:
+        super().__init__(scale)
+
+
+class LlamaCacheFlowAttention(GPTCacheFlowAttention):
+    """Llama uses GPT-NeoX style rotary embedding."""
+
+    def __init__(
+        self,
+        scale: float,
+        head_size: int,
+        max_position: int = 8192,
+        base: int = 10000,
+    ) -> None:
+        super().__init__(scale)
+
+        # Create the cos and sin cache.
+        inv_freq = 1.0 / (base ** (torch.arange(0, head_size, 2) / head_size))
+        t = torch.arange(max_position).float()
+        freqs = torch.einsum('i,j -> ij', t, inv_freq.float())
+        cos = freqs.cos()
+        sin = freqs.sin()
+        cache = torch.cat((cos, sin), dim=-1)
+
+        # FIXME(woosuk): This assumes that we configure the default dtype when
+        # initializing the model. Make it more robust.
+        torch_dtype = torch.get_default_dtype()
+        cache = cache.to(torch_dtype)
+        # Embedding size: [max_position, head_size]
+        self.register_buffer('cos_sin_cache', cache, persistent=False)
+
+    def forward(
+        self,
+        positions: torch.LongTensor,            # [num_tokens]
+        query: torch.Tensor,                    # [num_tokens, num_heads * head_size]
+        key: torch.Tensor,                      # [num_tokens, num_heads * head_size]
+        value: torch.Tensor,                    # [num_tokens, num_heads * head_size]
+        key_cache: torch.Tensor,                # [num_blocks, num_heads, head_size/x, block_size, x]
+        value_cache: torch.Tensor,              # [num_blocks, num_heads, head_size, block_size]
+        input_metadata: InputMetadata,
+        cache_event: Optional[torch.cuda.Event],
+    ) -> torch.Tensor:                          # [num_tokens, num_heads * head_size]
+        # Apply rotary embedding to the query and key before passing them
+        # to the attention op.
+        out_query = torch.empty_like(query)
+        out_key = torch.empty_like(key)
+        pos_encoding_ops.rotary_embedding_neox(
+            out_query,
+            out_key,
+            positions,
+            query,
+            key,
+            self.cos_sin_cache,
+        )
+        return super().forward(
+            out_query,
+            out_key,
+            value,
+            key_cache,
+            value_cache,
+            input_metadata,
+            cache_event,
+        )

cacheflow/models/llama.py

@@ -8,12 +8,10 @@
 import numpy as np
 import torch
 from torch import nn
-import torch.nn.functional as F
 from transformers import LlamaConfig
-from transformers import PreTrainedModel
 
 from cacheflow.models import InputMetadata
-from cacheflow.models.attention import OPTCacheFlowAttention
+from cacheflow.models.attention import LlamaCacheFlowAttention
 from cacheflow.models.sample import Sampler
 from cacheflow.parallel_utils.parallel_state import (
     get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
@@ -41,48 +39,8 @@ def forward(self, hidden_states):
         return self.weight * hidden_states
 
 
-class LlamaRotaryEmbedding(torch.nn.Module):
-
-    def __init__(self, dim, max_position_embeddings=2048, base=10000):
-        super().__init__()
-        self.max_position_embeddings = max_position_embeddings
-
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2) / dim))
-        self.register_buffer("inv_freq", inv_freq)
-
-        # Create cos and sin embeddings.
-        t = torch.arange(max_position_embeddings).float()
-        freqs = torch.einsum("i,j->ij", t, self.inv_freq.float())
-        emb = torch.cat((freqs, freqs), dim=-1)
-        cos = emb.cos().to(dtype=self.inv_freq.dtype)
-        sin = emb.sin().to(dtype=self.inv_freq.dtype)
-        self.register_buffer("cos_cached", cos, persistent=False)
-        self.register_buffer("sin_cached", sin, persistent=False)
-
-    def forward(
-        self,
-        positions: torch.LongTensor,
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        cos = F.embedding(positions, self.cos_cached)
-        sin = F.embedding(positions, self.sin_cached)
-        return cos, sin
-
-
-def rotate_half(x):
-    """Rotates half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(q, k, cos, sin):
-    # TODO: Optimize.
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-    return q_embed, k_embed
-
-
 class LlamaMLP(nn.Module):
+
     def __init__(
         self,
         hidden_size: int,
@@ -156,9 +114,7 @@ def __init__(
             input_is_parallel=True,
             perform_initialization=False,
         )
-        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim)
-        # FIXME(woosuk): Rename this.
-        self.attn = OPTCacheFlowAttention(scale=self.scaling)
+        self.attn = LlamaCacheFlowAttention(self.scaling, self.head_dim)
 
     def forward(
         self,
@@ -171,19 +127,9 @@ def forward(
         q, _ = self.q_proj(hidden_states)
         k, _ = self.k_proj(hidden_states)
         v, _ = self.v_proj(hidden_states)
-
-        # Apply rotrary embedding.
-        # TODO: Optimize.
-        q = q.view(-1, self.num_heads, self.head_dim).transpose(0, 1)
-        k = k.view(-1, self.num_heads, self.head_dim).transpose(0, 1)
-        cos, sin = self.rotary_emb(positions)
-        q, k = apply_rotary_pos_emb(q, k, cos, sin)
-        q = q.transpose(0, 1).contiguous().view(-1, self.num_heads * self.head_dim)
-        k = k.transpose(0, 1).contiguous().view(-1, self.num_heads * self.head_dim)
-
-        key_cache, value_cache = kv_cache
+        k_cache, v_cache = kv_cache
         attn_output = self.attn(
-            q, k, v, key_cache, value_cache, input_metadata, cache_event)
+            positions, q, k, v, k_cache, v_cache, input_metadata, cache_event)
         output, _ = self.o_proj(attn_output)
         return output
 

cacheflow/models/memory_analyzer.py

@@ -165,8 +165,7 @@ def __init__(
         self.head_size = config.hidden_size // self.num_heads
         self.ffn_size = config.intermediate_size
         self.vocab_size = config.vocab_size
-        # FIXME
-        self.max_position = 2048
+        self.max_position = 8192
 
     def _get_param_size(self) -> int:
         word_embedding = self.vocab_size * self.hidden_size // self.tensor_parallel_size

cacheflow/models/opt.py

@@ -51,7 +51,7 @@ def __init__(
         assert num_heads % tensor_model_parallel_world_size == 0
         self.num_heads = total_num_heads // tensor_model_parallel_world_size
         self.head_dim = embed_dim // total_num_heads
-        self.scaling = self.head_dim**-0.5
+        self.scaling = self.head_dim ** -0.5
 
         # TODO(woosuk): Fuse the three linear layers into one QKV linear layer.
         self.k_proj = ColumnParallelLinear(embed_dim, embed_dim, bias=bias,
@@ -66,7 +66,6 @@ def __init__(
         self.out_proj = RowParallelLinear(embed_dim, embed_dim, bias=bias,
                                           input_is_parallel=True,
                                           perform_initialization=False)
-
         self.attn = OPTCacheFlowAttention(scale=self.scaling)
 
     def forward(

cacheflow/models/sample.py

@@ -12,7 +12,7 @@
 class Sampler(nn.Module):
 
     def __init__(self) -> None:
-        super(Sampler, self).__init__()
+        super().__init__()
 
     def forward(
         self,

csrc/cache_kernels.cu

@@ -122,13 +122,13 @@ void reshape_and_cache(
   torch::Tensor& value_cache,
   torch::Tensor& slot_mapping) {
   int num_tokens = key.size(0);
-  int head_num = key.size(1);
+  int num_heads = key.size(1);
   int head_size = key.size(2);
   int block_size = key_cache.size(3);
   int x = key_cache.size(4);
 
   dim3 grid(num_tokens);
-  dim3 block(std::min(head_num * head_size, 512));
+  dim3 block(std::min(num_heads * head_size, 512));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
     key.scalar_type(),
@@ -140,7 +140,7 @@ void reshape_and_cache(
         key_cache.data_ptr<scalar_t>(),
         value_cache.data_ptr<scalar_t>(),
         slot_mapping.data_ptr<int>(),
-        head_num,
+        num_heads,
         head_size,
         block_size,
         x);

csrc/pos_encoding.cpp

@@ -0,0 +1,16 @@
+#include <torch/extension.h>
+
+void rotary_embedding_neox(
+  torch::Tensor& out_query,
+  torch::Tensor& out_key,
+  torch::Tensor& positions,
+  torch::Tensor& query,
+  torch::Tensor& key,
+  torch::Tensor& cos_sin_cache);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def(
+    "rotary_embedding_neox",
+    &rotary_embedding_neox,
+    "Apply GPT-NeoX style rotary embedding to query and key");
+}

csrc/pos_encoding_kernels.cu

@@ -0,0 +1,83 @@
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+
+namespace cacheflow {
+
+template<typename scalar_t>
+__global__ void rotary_embedding_neox_kernel(
+  scalar_t* __restrict__ out_query,             // [num_tokens, num_heads, head_size]
+  scalar_t* __restrict__ out_key,               // [num_tokens, num_heads, head_size]
+  const int64_t* __restrict__ positions,        // [num_tokens]
+  const scalar_t* __restrict__ query,           // [num_tokens, num_heads, head_size]
+  const scalar_t* __restrict__ key,             // [num_tokens, num_heads, head_size]
+  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, head_size // 2]
+  const int num_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 * head_size;
+
+  const int embed_dim = head_size / 2;
+  const int n = num_heads * head_size;
+  for (int i = threadIdx.x; i < n; i += blockDim.x) {
+    const int idx = token_idx * n + i;
+
+    const int head_idx = i / head_size;
+    const int head_offset = i % head_size;
+    const int token_head = token_idx * n + head_idx * head_size;
+
+    const bool is_first_half = head_offset < embed_dim;
+    const int rot_offset = head_offset % embed_dim;
+    const int x_index = rot_offset;
+    const int y_index = embed_dim + rot_offset;
+
+    const scalar_t cos = __ldg(cache_ptr + x_index);
+    const scalar_t sin = __ldg(cache_ptr + y_index);
+
+    const scalar_t q_x = __ldg(query + token_head + x_index);
+    const scalar_t q_y = __ldg(query + token_head + y_index);
+    const scalar_t q_cos = is_first_half ? q_x : q_y;
+    const scalar_t q_sin = is_first_half ? -q_y : q_x;
+    out_query[idx] = q_cos * cos + q_sin * sin;
+
+    const scalar_t k_x = __ldg(key + token_head + x_index);
+    const scalar_t k_y = __ldg(key + token_head + y_index);
+    const scalar_t k_cos = is_first_half ? k_x : k_y;
+    const scalar_t k_sin = is_first_half ? -k_y : k_x;
+    out_key[idx] = k_cos * cos + k_sin * sin;
+  }
+}
+
+} // namespace cacheflow
+
+void rotary_embedding_neox(
+  torch::Tensor& out_query,         // [num_tokens, num_heads * head_size]
+  torch::Tensor& out_key,           // [num_tokens, num_heads * head_size]
+  torch::Tensor& positions,         // [num_tokens]
+  torch::Tensor& query,             // [num_tokens, num_heads * head_size]
+  torch::Tensor& key,               // [num_tokens, num_heads * head_size]
+  torch::Tensor& cos_sin_cache)     // [max_position, head_size]
+{
+  int num_tokens = query.size(0);
+  int head_size = cos_sin_cache.size(1);
+  int num_heads = query.size(1) / head_size;
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(num_heads * head_size, 512));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+    query.scalar_type(),
+    "rotary_embedding_neox",
+    [&] {
+      cacheflow::rotary_embedding_neox_kernel<scalar_t><<<grid, block, 0, stream>>>(
+        out_query.data_ptr<scalar_t>(),
+        out_key.data_ptr<scalar_t>(),
+        positions.data_ptr<int64_t>(),
+        query.data_ptr<scalar_t>(),
+        key.data_ptr<scalar_t>(),
+        cos_sin_cache.data_ptr<scalar_t>(),
+        num_heads,
+        head_size);
+    });
+}

setup.py

@@ -23,6 +23,14 @@
 )
 ext_modules.append(attention_extension)
 
+# Positional encodings.
+positional_encoding_extension = cpp_extension.CUDAExtension(
+    name='cacheflow.pos_encoding_ops',
+    sources=['csrc/pos_encoding.cpp', 'csrc/pos_encoding_kernels.cu'],
+    extra_compile_args={'cxx': CXX_FLAGS, 'nvcc': NVCC_FLAGS},
+)
+ext_modules.append(positional_encoding_extension)
+
 setuptools.setup(
     name='cacheflow',
     ext_modules=ext_modules,