Enable GQA support in the prefix prefill kernels (#3007)

Signed-off-by: Tao He <sighingnow@gmail.com>
This commit is contained in:
Tao He
2024-02-27 01:14:31 -08:00
committed by GitHub
parent 8b430d7dea
commit 71bcaf99e2
3 changed files with 87 additions and 47 deletions
+18 -16
View File
@@ -137,25 +137,27 @@ class PagedAttention(nn.Module):
)
if input_metadata.is_prompt:
# Prompt run.
if self.num_kv_heads != self.num_heads:
# As of Nov 2023, xformers only supports MHA. For MQA/GQA,
# project the key and value tensors to the desired number of
# heads.
# TODO(woosuk): Use MQA/GQA kernels for higher performance.
query = query.view(query.shape[0], self.num_kv_heads,
self.num_queries_per_kv, query.shape[-1])
key = key[:, :,
None, :].expand(key.shape[0], self.num_kv_heads,
self.num_queries_per_kv,
key.shape[-1])
value = value[:, :, None, :].expand(value.shape[0],
self.num_kv_heads,
self.num_queries_per_kv,
value.shape[-1])
# normal attention
if (key_cache is None or value_cache is None
or input_metadata.block_tables.numel() == 0):
if self.num_kv_heads != self.num_heads:
# As of Nov 2023, xformers only supports MHA. For MQA/GQA,
# project the key and value tensors to the desired number of
# heads.
# TODO(woosuk): Use MQA/GQA kernels for higher performance.
query = query.view(query.shape[0], self.num_kv_heads,
self.num_queries_per_kv,
query.shape[-1])
key = key[:, :,
None, :].expand(key.shape[0], self.num_kv_heads,
self.num_queries_per_kv,
key.shape[-1])
value = value[:, :,
None, :].expand(value.shape[0],
self.num_kv_heads,
self.num_queries_per_kv,
value.shape[-1])
# Set attention bias if not provided. This typically happens at
# the very attention layer of every iteration.
# FIXME(woosuk): This is a hack.
@@ -45,6 +45,7 @@ if triton.__version__ >= "2.1.0":
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
@@ -53,6 +54,8 @@ if triton.__version__ >= "2.1.0":
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
@@ -85,13 +88,14 @@ if triton.__version__ >= "2.1.0":
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_head * stride_k_cache_h +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs + cur_head * stride_v_cache_h +
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
@@ -131,9 +135,9 @@ if triton.__version__ >= "2.1.0":
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_head * stride_kh +
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_head * stride_vh +
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
@@ -232,6 +236,7 @@ if triton.__version__ >= "2.1.0":
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
@@ -240,6 +245,8 @@ if triton.__version__ >= "2.1.0":
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
@@ -272,13 +279,14 @@ if triton.__version__ >= "2.1.0":
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_head * stride_k_cache_h +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs + cur_head * stride_v_cache_h +
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
@@ -317,9 +325,9 @@ if triton.__version__ >= "2.1.0":
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_head * stride_kh +
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_head * stride_vh +
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
@@ -420,6 +428,7 @@ if triton.__version__ >= "2.1.0":
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
@@ -429,6 +438,8 @@ if triton.__version__ >= "2.1.0":
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
# cur_batch_seq_len: the length of prompts
# cur_batch_ctx_len: the length of prefix
# cur_batch_in_all_start_index: the start id of the dim=0
@@ -468,13 +479,14 @@ if triton.__version__ >= "2.1.0":
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_head * stride_k_cache_h +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs + cur_head * stride_v_cache_h +
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
@@ -522,9 +534,9 @@ if triton.__version__ >= "2.1.0":
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_head * stride_kh +
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_head * stride_vh +
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
@@ -628,6 +640,7 @@ if triton.__version__ >= "2.1.0":
sm_scale = 1.0 / (Lq**0.5)
batch, head = b_seq_len.shape[0], q.shape[1]
num_queries_per_kv = q.shape[1] // k.shape[1]
grid = (batch, head, triton.cdiv(max_input_len, BLOCK)) # batch, head,
@@ -674,6 +687,7 @@ if triton.__version__ >= "2.1.0":
v_cache.stride(2),
v_cache.stride(
3), #[num_blocks, num_kv_heads, head_size, block_size]
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_N=BLOCK,
@@ -721,6 +735,7 @@ if triton.__version__ >= "2.1.0":
v_cache.stride(2),
v_cache.stride(
3), #[num_blocks, num_kv_heads, head_size, block_size]
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_N=BLOCK,