xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 09:35:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Kernel] Chunk-aligned mamba2 (#24683)

Browse Source
This commit is contained in:
Thomas Parnell 2025-09-29 23:18:25 +02:00 committed by GitHub
parent 61a3431613
commit fea3e476aa
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 247 additions and 431 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/layers/mamba/mamba_mixer2.py
View File

@ -502,9 +502,9 @@ class MambaMixer2(MambaBase, CustomOp):
prep_initial_states = attn_metadata.prep_initial_states prep_initial_states = attn_metadata.prep_initial_states
chunk_size = attn_metadata.chunk_size chunk_size = attn_metadata.chunk_size
seq_idx_p = attn_metadata.seq_idx_p seq_idx_p = attn_metadata.seq_idx_p
chunk_indices_p = attn_metadata.chunk_indices_p
chunk_offsets_p = attn_metadata.chunk_offsets_p
query_start_loc_p = attn_metadata.query_start_loc_p query_start_loc_p = attn_metadata.query_start_loc_p
cu_chunk_seqlen_p = attn_metadata.cu_chunk_seqlen_p
last_chunk_indices_p = attn_metadata.last_chunk_indices_p
# 1. Gated MLP's linear projection # 1. Gated MLP's linear projection
projected_states, _ = self.in_proj(hidden_states) projected_states, _ = self.in_proj(hidden_states)
@ -634,9 +634,9 @@ class MambaMixer2(MambaBase, CustomOp):
z=None, z=None,
dt_bias=self.dt_bias, dt_bias=self.dt_bias,
seq_idx=seq_idx_p, seq_idx=seq_idx_p,
chunk_indices=chunk_indices_p,
chunk_offsets=chunk_offsets_p,
cu_seqlens=query_start_loc_p, cu_seqlens=query_start_loc_p,
cu_chunk_seqlens=cu_chunk_seqlen_p,
last_chunk_indices=last_chunk_indices_p,
initial_states=initial_states, initial_states=initial_states,
dt_softplus=True, dt_softplus=True,
dt_limit=(0.0, float("inf")), dt_limit=(0.0, float("inf")),

42
vllm/model_executor/layers/mamba/ops/ssd_bmm.py
View File

@ -6,8 +6,6 @@
# ruff: noqa: E501,SIM102 # ruff: noqa: E501,SIM102
import math
import torch import torch
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
@ -96,7 +94,7 @@ def _bmm_chunk_fwd_kernel(
a_ptr, a_ptr,
b_ptr, b_ptr,
out_ptr, out_ptr,
seq_idx_ptr, cu_chunk_seqlens_ptr,
# Matrix dimensions # Matrix dimensions
seqlen, seqlen,
chunk_size: tl.constexpr, chunk_size: tl.constexpr,
@ -112,7 +110,6 @@ def _bmm_chunk_fwd_kernel(
stride_out_head: tl.int64, stride_out_head: tl.int64,
stride_outm: tl.int64, stride_outm: tl.int64,
stride_outn: tl.constexpr, stride_outn: tl.constexpr,
stride_seq_idx_seqlen: tl.constexpr,
# Meta-parameters # Meta-parameters
IS_CAUSAL: tl.constexpr, IS_CAUSAL: tl.constexpr,
dot_dtype: tl.constexpr, dot_dtype: tl.constexpr,
@ -129,10 +126,12 @@ def _bmm_chunk_fwd_kernel(
if IS_CAUSAL: if IS_CAUSAL:
if pid_n * BLOCK_SIZE_N >= (pid_m + 1) * BLOCK_SIZE_M: if pid_n * BLOCK_SIZE_N >= (pid_m + 1) * BLOCK_SIZE_M:
return return
a_ptr += pid_c * chunk_size * stride_a_seqlen + pid_h * stride_a_head
b_ptr += pid_c * chunk_size * stride_b_seqlen + pid_h * stride_b_head
seq_idx_ptr += pid_c * chunk_size * stride_seq_idx_seqlen chunk_seqlen_start = tl.load(cu_chunk_seqlens_ptr + pid_c)
chunk_seqlen_end = tl.load(cu_chunk_seqlens_ptr + pid_c + 1)
a_ptr += chunk_seqlen_start * stride_a_seqlen + pid_h * stride_a_head
b_ptr += chunk_seqlen_start * stride_b_seqlen + pid_h * stride_b_head
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
@ -141,7 +140,7 @@ def _bmm_chunk_fwd_kernel(
offs_k[None, :] * stride_ak) offs_k[None, :] * stride_ak)
b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + b_ptrs = b_ptr + (offs_k[:, None] * stride_bk +
offs_n[None, :] * stride_b_seqlen) offs_n[None, :] * stride_b_seqlen)
chunk_size_limit = min(chunk_size, seqlen - pid_c * chunk_size) chunk_size_limit = chunk_seqlen_end - chunk_seqlen_start
acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
@ -162,16 +161,6 @@ def _bmm_chunk_fwd_kernel(
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
# Zero out the results that are not from the same request
# in the varlen batch
seq_idx_m = tl.load(seq_idx_ptr + offs_m * stride_seq_idx_seqlen,
mask=offs_m < chunk_size_limit,
other=-1)
seq_idx_n = tl.load(seq_idx_ptr + offs_n * stride_seq_idx_seqlen,
mask=offs_n < chunk_size_limit,
other=-2)
acc = tl.where(seq_idx_m[:, None] == seq_idx_n[None, :], acc, 0.0)
out = acc.to(out_ptr.dtype.element_ty) out = acc.to(out_ptr.dtype.element_ty)
out_ptr += pid_c * stride_out_chunk + pid_h * stride_out_head out_ptr += pid_c * stride_out_chunk + pid_h * stride_out_head
out_ptrs = out_ptr + (stride_outm * offs_m[:, None] + out_ptrs = out_ptr + (stride_outm * offs_m[:, None] +
@ -182,12 +171,18 @@ def _bmm_chunk_fwd_kernel(
(offs_n[None, :] < chunk_size)) (offs_n[None, :] < chunk_size))
def _bmm_chunk_fwd(a, b, chunk_size, seq_idx, causal=False, output_dtype=None): def _bmm_chunk_fwd(a,
b,
chunk_size,
cu_chunk_seqlens,
causal=False,
output_dtype=None):
""" """
Argument: Argument:
a: (seqlen, ngroups, k) a: (seqlen, ngroups, k)
b: (seqlen, ngroups, k) b: (seqlen, ngroups, k)
seq_idx: (seqlen,). out[i, j] for seq_idx[i] != seq_idx[j] will be zeroed out. chunk_size: int
cu_chunk_seq_lens: (nchunks+1,)
causal: if True, then out[i, j] for i > j will be arbitrary, only out[i, j] for i <= j are causal: if True, then out[i, j] for i > j will be arbitrary, only out[i, j] for i <= j are
guaranteed to be correct. guaranteed to be correct.
Return: Return:
@ -195,14 +190,12 @@ def _bmm_chunk_fwd(a, b, chunk_size, seq_idx, causal=False, output_dtype=None):
""" """
seqlen, ngroups, k = a.shape seqlen, ngroups, k = a.shape
assert b.shape == a.shape assert b.shape == a.shape
assert seq_idx is not None
assert seq_idx.shape == (seqlen, )
if a.stride(-1) != 1 and a.stride(0) != 1: if a.stride(-1) != 1 and a.stride(0) != 1:
a = a.contiguous() a = a.contiguous()
if b.stride(-1) != 1 and b.stride(0) != 1: if b.stride(-1) != 1 and b.stride(0) != 1:
b = b.contiguous() b = b.contiguous()
nchunks = math.ceil(seqlen / chunk_size) nchunks = len(cu_chunk_seqlens) - 1
# Allocates output. # Allocates output.
out_dtype = a.dtype if output_dtype is None else output_dtype out_dtype = a.dtype if output_dtype is None else output_dtype
out = torch.empty((nchunks, ngroups, chunk_size, chunk_size), out = torch.empty((nchunks, ngroups, chunk_size, chunk_size),
@ -220,7 +213,7 @@ def _bmm_chunk_fwd(a, b, chunk_size, seq_idx, causal=False, output_dtype=None):
a_ptr=a, a_ptr=a,
b_ptr=b, b_ptr=b,
out_ptr=out, out_ptr=out,
seq_idx_ptr=seq_idx, cu_chunk_seqlens_ptr=cu_chunk_seqlens,
seqlen=seqlen, seqlen=seqlen,
chunk_size=chunk_size, chunk_size=chunk_size,
K=k, K=k,
@ -235,7 +228,6 @@ def _bmm_chunk_fwd(a, b, chunk_size, seq_idx, causal=False, output_dtype=None):
stride_out_head=out.stride(1), stride_out_head=out.stride(1),
stride_outm=out.stride(-2), stride_outm=out.stride(-2),
stride_outn=out.stride(-1), stride_outn=out.stride(-1),
stride_seq_idx_seqlen=seq_idx.stride(0),
IS_CAUSAL=causal, IS_CAUSAL=causal,
dot_dtype=dot_dtype, dot_dtype=dot_dtype,
) )

232
vllm/model_executor/layers/mamba/ops/ssd_chunk_scan.py
View File

@ -120,9 +120,7 @@ def _chunk_scan_fwd_kernel(
states_ptr, states_ptr,
D_ptr, D_ptr,
initstates_ptr, initstates_ptr,
chunk_indices_ptr, cu_chunk_seqlens_ptr,
chunk_offsets_ptr,
chunk_meta_num,
# Matrix dimensions # Matrix dimensions
chunk_size: tl.constexpr, chunk_size: tl.constexpr,
hdim: tl.constexpr, hdim: tl.constexpr,
@ -149,7 +147,7 @@ def _chunk_scan_fwd_kernel(
stride_dA_cs_chunk: tl.int64, stride_dA_cs_chunk: tl.int64,
stride_dA_cs_head: tl.int64, stride_dA_cs_head: tl.int64,
stride_dA_cs_csize: tl.constexpr, stride_dA_cs_csize: tl.constexpr,
stride_seq_idx_seqlen: tl.constexpr, stride_seq_idx_chunk: tl.constexpr,
stride_C_seqlen: tl.int64, stride_C_seqlen: tl.int64,
stride_C_head: tl.int64, stride_C_head: tl.int64,
stride_C_dstate: tl.constexpr, stride_C_dstate: tl.constexpr,
@ -175,170 +173,107 @@ def _chunk_scan_fwd_kernel(
HAS_INITSTATES: tl.constexpr, HAS_INITSTATES: tl.constexpr,
): ):
pid_c = tl.program_id(axis=1).to(tl.int64) pid_c = tl.program_id(axis=1).to(tl.int64)
if not HAS_INITSTATES:
c_idx = pid_c
c_off = 0
else:
c_idx = tl.load(chunk_indices_ptr + pid_c, mask=pid_c > -1, other=0)
c_off = tl.load(chunk_offsets_ptr + pid_c, mask=pid_c > -1, other=0)
pid_h = tl.program_id(axis=2) pid_h = tl.program_id(axis=2)
num_pid_n = tl.cdiv(hdim, BLOCK_SIZE_N) num_pid_n = tl.cdiv(hdim, BLOCK_SIZE_N)
pid_m = tl.program_id(axis=0) // num_pid_n pid_m = tl.program_id(axis=0) // num_pid_n
pid_n = tl.program_id(axis=0) % num_pid_n pid_n = tl.program_id(axis=0) % num_pid_n
cb_ptr += c_idx * stride_cb_chunk + (pid_h // cb_ptr += pid_c * stride_cb_chunk + (pid_h //
nheads_ngroups_ratio) * stride_cb_head nheads_ngroups_ratio) * stride_cb_head
x_ptr += c_idx * chunk_size * stride_x_seqlen + pid_h * stride_x_head chunk_seqlen_start = tl.load(cu_chunk_seqlens_ptr + pid_c)
dt_ptr += c_idx * stride_dt_chunk + pid_h * stride_dt_head chunk_seqlen_end = tl.load(cu_chunk_seqlens_ptr + pid_c + 1)
dA_cumsum_ptr += c_idx * stride_dA_cs_chunk + pid_h * stride_dA_cs_head x_ptr += chunk_seqlen_start * stride_x_seqlen + pid_h * stride_x_head
C_ptr += c_idx * chunk_size * stride_C_seqlen + ( dt_ptr += pid_c * stride_dt_chunk + pid_h * stride_dt_head
dA_cumsum_ptr += pid_c * stride_dA_cs_chunk + pid_h * stride_dA_cs_head
C_ptr += chunk_seqlen_start * stride_C_seqlen + (
pid_h // nheads_ngroups_ratio) * stride_C_head pid_h // nheads_ngroups_ratio) * stride_C_head
# M-block offsets and prev states # M-block offsets and prev states
# - logic in next block may override these if there is an active offset # - logic in next block may override these if there is an active offset
offs_m = pid_m * BLOCK_SIZE_M + c_off + tl.arange(0, BLOCK_SIZE_M) offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
prev_states_ptr = states_ptr + c_idx * stride_states_chunk + pid_h * stride_states_head
prev_states_hdim = stride_states_hdim
prev_states_dstate = stride_states_dstate
chunk_size_limit = min(chunk_size, seqlen - c_idx * chunk_size) seq_idx_ptr += pid_c * stride_seq_idx_chunk
seq_idx = tl.load(seq_idx_ptr)
seq_idx_prev = tl.load(seq_idx_ptr - stride_seq_idx_chunk,
mask=pid_c >= 1,
other=-1)
seq_idx_ptr += c_idx * chunk_size * stride_seq_idx_seqlen if HAS_INITSTATES and (seq_idx != seq_idx_prev):
# - we only need seq_idx_prev to be aligned to chunk boundary prev_states_ptr = initstates_ptr + seq_idx * stride_init_states_batch + pid_h * stride_init_states_head
seq_idx_prev = tl.load(seq_idx_ptr - stride_seq_idx_seqlen, prev_states_hdim = stride_init_states_hdim
mask=c_idx >= 1, prev_states_dstate = stride_init_states_dstate
other=0) else:
prev_states_ptr = states_ptr + (
pid_c - 1) * stride_states_chunk + pid_h * stride_states_head
prev_states_hdim = stride_states_hdim
prev_states_dstate = stride_states_dstate
if HAS_INITSTATES: chunk_size_limit = chunk_seqlen_end - chunk_seqlen_start
# if there are init states, we only need seq_idx_m to point
# what is the current seq_idx
# get current seq idx
if (pid_m * BLOCK_SIZE_M + c_off) < chunk_size_limit:
seq_idx_m = tl.load(
seq_idx_ptr +
(pid_m * BLOCK_SIZE_M + c_off) * stride_seq_idx_seqlen, )
# - recall that in ssd_state_passing, for the case c_off == 0
# i.e., the very first sequence, we made states_ptr hold its initial state
# so this edge case is taken care of
if ((c_off == 0) and (seq_idx_prev != seq_idx_m
) # if a seq is changed exactly on boundary
or (c_off > 0) # implies a new example (pseudo chunk)
):
# - replace prev_states_ptr with init_states
prev_states_ptr = initstates_ptr + seq_idx_m * stride_init_states_batch + pid_h * stride_init_states_head
prev_states_hdim = stride_init_states_hdim # override strides
prev_states_dstate = stride_init_states_dstate
offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
dA_cs_m = tl.load(dA_cumsum_ptr + offs_m * stride_dA_cs_csize, dA_cs_m = tl.load(dA_cumsum_ptr + offs_m * stride_dA_cs_csize,
mask=offs_m < chunk_size, mask=offs_m < chunk_size,
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
# - handle chunk state limit
if HAS_INITSTATES:
# have to split this if otherwise compilation will have problems
dA_cs_m_boundary = 0.0
# get the c_idx for the next (logica) chunk
c_idx_n = tl.load(
chunk_indices_ptr + (pid_c + 1),
mask=pid_c > -1 and (pid_c + 1) < chunk_meta_num,
other=-1 # to trigger different chunk
)
# - there are things to consider
# A. if c_off > 0 then we need to move the dA_cs boundary to ensure correct
# contribution of past states
# B. if c_off_n < chunk_size_limit, then we need to adjust this so as not to
# encroach into the next sequence, where c_off_n is the offset of the next
# (logical) chunk.
# An equivalent check for B is c_idx == c_idx_n, where there is repetition in
# (logical) chunk indices.
if (c_idx == c_idx_n) or c_off > 0:
# get the next offset
c_off_n = tl.load(chunk_offsets_ptr + (pid_c + 1),
mask=pid_c > -1 and (pid_c + 1) < chunk_meta_num,
other=chunk_size)
# in this case, adjust down the chunk_size_limit
if c_idx == c_idx_n:
chunk_size_limit = min(c_off_n, chunk_size_limit)
# get the cs at the offset boundary
# - c_off == 0 is a passthrough
# - We need dA_cs at the boundary, defined by c_off - no need
# to increase pointer by pid_m (it is a constant offset,
# i.e. the same for all blocks)
dA_cs_m_boundary = tl.load(
dA_cumsum_ptr + (c_off - 1) * stride_dA_cs_csize,
mask=(((c_off - 1) > -1) and ((c_off) < chunk_size)),
other=0.0).to(tl.float32)
else:
# - handle seq idx when HAS_INITSTATES==False
seq_idx_m = tl.load(seq_idx_ptr + offs_m * stride_seq_idx_seqlen,
mask=offs_m < chunk_size_limit,
other=-1)
acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
# Without the if (pid_c > -1), with Triton 2.1.0, I get offs_out_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
# Assertion `!(srcMmaLayout && dstMmaLayout) && "Unexpected mma -> mm a layout conversion"' failed. offs_out_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
# With Triton 2.2.0, this works
if IS_TRITON_22 or c_idx > -1:
# Faster to just do 1 iteration with larger BLOCK_SIZE_K, up to block size 128
offs_k_dstate = tl.arange(
0, BLOCK_SIZE_DSTATE if BLOCK_SIZE_DSTATE <= 128 else BLOCK_SIZE_K)
C_ptrs = C_ptr + (offs_m[:, None] * stride_C_seqlen +
offs_k_dstate[None, :] * stride_C_dstate)
prev_states_ptrs = prev_states_ptr + ( # Faster to just do 1 iteration with larger BLOCK_SIZE_K, up to block size 128
offs_n[None, :] * prev_states_hdim + offs_k_dstate = tl.arange(
offs_k_dstate[:, None] * prev_states_dstate) 0, BLOCK_SIZE_DSTATE if BLOCK_SIZE_DSTATE <= 128 else BLOCK_SIZE_K)
C_ptrs = C_ptr + (offs_m[:, None] * stride_C_seqlen +
offs_k_dstate[None, :] * stride_C_dstate)
if not HAS_INITSTATES: scale_m = tl.exp(dA_cs_m)
# - this is for continuous batching where there is no init states if BLOCK_SIZE_DSTATE <= 128:
scale_m = tl.where(seq_idx_m == seq_idx_prev, tl.exp(dA_cs_m), 0.0) C = tl.load(C_ptrs,
mask=(offs_m[:, None] < chunk_size_limit) &
(offs_k_dstate[None, :] < dstate),
other=0.0)
if not HAS_INITSTATES and (seq_idx != seq_idx_prev):
# if no init states AND starting a new sequence, we need zeros
prev_states = tl.zeros((BLOCK_SIZE_DSTATE, BLOCK_SIZE_N),
dtype=C_ptr.dtype.element_ty)
else: else:
# - if there is initstates, we will rely on prev_states, no zeroing # otherwise read the previous state
# required. prev_states_ptrs = prev_states_ptr \
scale_m = tl.exp(dA_cs_m - dA_cs_m_boundary) + offs_n[None, :] * prev_states_hdim \
+ offs_k_dstate[:, None] * prev_states_dstate
if BLOCK_SIZE_DSTATE <= 128:
C = tl.load(C_ptrs,
mask=(offs_m[:, None] < chunk_size_limit) &
(offs_k_dstate[None, :] < dstate),
other=0.0)
prev_states = tl.load(prev_states_ptrs, prev_states = tl.load(prev_states_ptrs,
mask=(offs_k_dstate[:, None] < dstate) & mask=(offs_k_dstate[:, None] < dstate) &
(offs_n[None, :] < hdim), (offs_n[None, :] < hdim),
other=0.0) other=0.0)
prev_states = prev_states.to(C_ptr.dtype.element_ty) prev_states = prev_states.to(C_ptr.dtype.element_ty)
acc = tl.dot(C, prev_states) * scale_m[:, None]
else: acc = tl.dot(C, prev_states) * scale_m[:, None]
for k in range(0, dstate, BLOCK_SIZE_K):
C = tl.load(C_ptrs, else:
mask=(offs_m[:, None] < chunk_size_limit) & prev_states_ptrs = prev_states_ptr \
(offs_k_dstate[None, :] < dstate - k), + offs_n[None, :] * prev_states_hdim \
other=0.0) + offs_k_dstate[:, None] * prev_states_dstate
# C = (C * scale_m[:, None]).to(C_ptr.dtype.element_ty) for k in range(0, dstate, BLOCK_SIZE_K):
C = tl.load(C_ptrs,
mask=(offs_m[:, None] < chunk_size_limit) &
(offs_k_dstate[None, :] < dstate - k),
other=0.0)
if not HAS_INITSTATES and (seq_idx != seq_idx_prev):
prev_states = tl.zeros((BLOCK_SIZE_DSTATE, BLOCK_SIZE_K),
dtype=C_ptr.dtype.element_ty)
else:
prev_states = tl.load( prev_states = tl.load(
prev_states_ptrs, prev_states_ptrs,
mask=(offs_k_dstate[:, None] < dstate - k) & mask=(offs_k_dstate[:, None] < dstate - k) &
(offs_n[None, :] < hdim), (offs_n[None, :] < hdim),
other=0.0) other=0.0)
prev_states = prev_states.to(C_ptr.dtype.element_ty) prev_states = prev_states.to(C_ptr.dtype.element_ty)
acc += tl.dot(C, prev_states) acc += tl.dot(C, prev_states)
C_ptrs += BLOCK_SIZE_K C_ptrs += BLOCK_SIZE_K
prev_states_ptrs += BLOCK_SIZE_K prev_states_ptrs += BLOCK_SIZE_K
acc *= scale_m[:, None] acc *= scale_m[:, None]
offs_k = tl.arange(0, BLOCK_SIZE_K) + c_off offs_k = tl.arange(0, BLOCK_SIZE_K)
cb_ptrs = cb_ptr + (offs_m[:, None] * stride_cb_csize_m + cb_ptrs = cb_ptr + (offs_m[:, None] * stride_cb_csize_m +
offs_k[None, :] * stride_cb_csize_k) offs_k[None, :] * stride_cb_csize_k)
x_ptrs = x_ptr + (offs_k[:, None] * stride_x_seqlen + x_ptrs = x_ptr + (offs_k[:, None] * stride_x_seqlen +
@ -375,7 +310,7 @@ def _chunk_scan_fwd_kernel(
dt_ptrs += BLOCK_SIZE_K * stride_dt_csize dt_ptrs += BLOCK_SIZE_K * stride_dt_csize
dA_cumsum_ptrs += BLOCK_SIZE_K * stride_dA_cs_csize dA_cumsum_ptrs += BLOCK_SIZE_K * stride_dA_cs_csize
offs_out_m = pid_m * BLOCK_SIZE_M + c_off + tl.arange(0, BLOCK_SIZE_M) offs_out_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_out_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) offs_out_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
if HAS_D: if HAS_D:
@ -393,7 +328,7 @@ def _chunk_scan_fwd_kernel(
acc += x_residual * D acc += x_residual * D
if HAS_Z: if HAS_Z:
z_ptr += c_idx * chunk_size * stride_z_seqlen + pid_h * stride_z_head z_ptr += chunk_seqlen_start * stride_z_seqlen + pid_h * stride_z_head
z_ptrs = z_ptr + (stride_z_seqlen * offs_out_m[:, None] + z_ptrs = z_ptr + (stride_z_seqlen * offs_out_m[:, None] +
stride_z_hdim * offs_out_n[None, :]) stride_z_hdim * offs_out_n[None, :])
z = tl.load(z_ptrs, z = tl.load(z_ptrs,
@ -402,7 +337,7 @@ def _chunk_scan_fwd_kernel(
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
acc *= z * tl.sigmoid(z) acc *= z * tl.sigmoid(z)
out_ptr += c_idx * chunk_size * stride_out_seqlen + pid_h * stride_out_head out_ptr += chunk_seqlen_start * stride_out_seqlen + pid_h * stride_out_head
out_ptrs = out_ptr + (stride_out_seqlen * offs_out_m[:, None] + out_ptrs = out_ptr + (stride_out_seqlen * offs_out_m[:, None] +
offs_out_n[None, :] * stride_out_hdim) offs_out_n[None, :] * stride_out_hdim)
tl.store(out_ptrs, tl.store(out_ptrs,
@ -418,12 +353,11 @@ def _chunk_scan_fwd(
dA_cumsum, dA_cumsum,
C, C,
states, states,
cu_chunk_seqlens,
out, out,
seq_idx, seq_idx,
D=None, D=None,
z=None, z=None,
chunk_indices=None,
chunk_offsets=None,
initial_states=None, initial_states=None,
): ):
assert seq_idx is not None, "this implementation requires seq_idx" assert seq_idx is not None, "this implementation requires seq_idx"
@ -441,20 +375,10 @@ def _chunk_scan_fwd(
assert dt.shape == (nheads, nchunks, chunk_size) assert dt.shape == (nheads, nchunks, chunk_size)
assert dA_cumsum.shape == (nheads, nchunks, chunk_size) assert dA_cumsum.shape == (nheads, nchunks, chunk_size)
assert states.shape == (nchunks, nheads, headdim, dstate) assert states.shape == (nchunks, nheads, headdim, dstate)
assert seq_idx.shape == (seqlen, ) assert seq_idx.shape == (nchunks, )
if initial_states is not None: grid = lambda META: (triton.cdiv(chunk_size, META['BLOCK_SIZE_M']) * triton
# with initial states, we need to take care of how .cdiv(headdim, META['BLOCK_SIZE_N']), nchunks, nheads)
# seq_idx crosses the boundaries
assert chunk_indices is not None and chunk_offsets is not None, \
"chunk_indices and chunk_offsets should have been set"
else:
chunk_indices, chunk_offsets = None, None
grid = lambda META: (
triton.cdiv(chunk_size, META['BLOCK_SIZE_M']) * triton.cdiv(
headdim, META['BLOCK_SIZE_N']), nchunks
if chunk_offsets is None else len(chunk_offsets), nheads)
z_strides = ((z.stride(0), z.stride(1), z.stride(2)) if z is not None else z_strides = ((z.stride(0), z.stride(1), z.stride(2)) if z is not None else
(0, 0, 0)) (0, 0, 0))
@ -476,9 +400,7 @@ def _chunk_scan_fwd(
states_ptr=states, states_ptr=states,
D_ptr=D, D_ptr=D,
initstates_ptr=initial_states, initstates_ptr=initial_states,
chunk_indices_ptr=chunk_indices, cu_chunk_seqlens_ptr=cu_chunk_seqlens,
chunk_offsets_ptr=chunk_offsets,
chunk_meta_num=len(chunk_indices) if chunk_indices is not None else 0,
chunk_size=chunk_size, chunk_size=chunk_size,
hdim=headdim, hdim=headdim,
dstate=dstate, dstate=dstate,
@ -503,7 +425,7 @@ def _chunk_scan_fwd(
stride_dA_cs_chunk=dA_cumsum.stride(1), stride_dA_cs_chunk=dA_cumsum.stride(1),
stride_dA_cs_head=dA_cumsum.stride(0), stride_dA_cs_head=dA_cumsum.stride(0),
stride_dA_cs_csize=dA_cumsum.stride(2), stride_dA_cs_csize=dA_cumsum.stride(2),
stride_seq_idx_seqlen=seq_idx.stride(0), stride_seq_idx_chunk=seq_idx.stride(0),
stride_C_seqlen=C.stride(0), stride_C_seqlen=C.stride(0),
stride_C_head=C.stride(1), stride_C_head=C.stride(1),
stride_C_dstate=C.stride(2), stride_C_dstate=C.stride(2),

48
vllm/model_executor/layers/mamba/ops/ssd_chunk_state.py
View File

@ -6,8 +6,6 @@
# ruff: noqa: E501 # ruff: noqa: E501
import math
import torch import torch
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
@ -34,6 +32,7 @@ def _chunk_cumsum_fwd_kernel(
dt_bias_ptr, dt_bias_ptr,
dt_out_ptr, dt_out_ptr,
dA_cumsum_ptr, dA_cumsum_ptr,
cu_chunk_seqlens_ptr,
# Matrix dimension # Matrix dimension
seqlen, seqlen,
nheads: tl.constexpr, nheads: tl.constexpr,
@ -61,7 +60,11 @@ def _chunk_cumsum_fwd_kernel(
# https://github.com/triton-lang/triton/issues/1058 # https://github.com/triton-lang/triton/issues/1058
pid_c = tl.program_id(axis=0).to(tl.int64) pid_c = tl.program_id(axis=0).to(tl.int64)
pid_h = tl.program_id(axis=1) pid_h = tl.program_id(axis=1)
dt_ptr += pid_c * chunk_size * stride_dt_seqlen
chunk_seqlen_start = tl.load(cu_chunk_seqlens_ptr + pid_c)
chunk_seqlen_end = tl.load(cu_chunk_seqlens_ptr + pid_c + 1)
dt_ptr += chunk_seqlen_start * stride_dt_seqlen
dt_out_ptr += pid_c * stride_dt_out_chunk dt_out_ptr += pid_c * stride_dt_out_chunk
dA_cumsum_ptr += pid_c * stride_dA_cs_chunk dA_cumsum_ptr += pid_c * stride_dA_cs_chunk
@ -74,7 +77,7 @@ def _chunk_cumsum_fwd_kernel(
offs_c[None, :] * stride_dt_out_csize) offs_c[None, :] * stride_dt_out_csize)
dA_cs_ptrs = dA_cumsum_ptr + (offs_h[:, None] * stride_dA_cs_head + dA_cs_ptrs = dA_cumsum_ptr + (offs_h[:, None] * stride_dA_cs_head +
offs_c[None, :] * stride_dA_cs_csize) offs_c[None, :] * stride_dA_cs_csize)
chunk_size_limit = min(chunk_size, seqlen - pid_c * chunk_size) chunk_size_limit = chunk_seqlen_end - chunk_seqlen_start
dt = tl.load(dt_ptrs, dt = tl.load(dt_ptrs,
mask=(offs_h[:, None] < nheads) & mask=(offs_h[:, None] < nheads) &
@ -188,7 +191,7 @@ def _chunk_state_fwd_kernel(
states_ptr, states_ptr,
dt_ptr, dt_ptr,
dA_cumsum_ptr, dA_cumsum_ptr,
seq_idx_ptr, cu_chunk_seqlens_ptr,
# Matrix dimensions # Matrix dimensions
hdim: tl.constexpr, hdim: tl.constexpr,
dstate: tl.constexpr, dstate: tl.constexpr,
@ -212,7 +215,6 @@ def _chunk_state_fwd_kernel(
stride_dA_cs_head: tl.int64, stride_dA_cs_head: tl.int64,
stride_dA_cs_chunk: tl.int64, stride_dA_cs_chunk: tl.int64,
stride_dA_cs_csize: tl.constexpr, stride_dA_cs_csize: tl.constexpr,
stride_seq_idx_seqlen: tl.constexpr,
# Meta-parameters # Meta-parameters
BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_N: tl.constexpr,
@ -223,14 +225,14 @@ def _chunk_state_fwd_kernel(
num_pid_n = tl.cdiv(dstate, BLOCK_SIZE_N) num_pid_n = tl.cdiv(dstate, BLOCK_SIZE_N)
pid_m = tl.program_id(axis=0) // num_pid_n pid_m = tl.program_id(axis=0) // num_pid_n
pid_n = tl.program_id(axis=0) % num_pid_n pid_n = tl.program_id(axis=0) % num_pid_n
b_ptr += pid_c * chunk_size * stride_b_seqlen + ( chunk_seqlen_start = tl.load(cu_chunk_seqlens_ptr + pid_c)
chunk_seqlen_end = tl.load(cu_chunk_seqlens_ptr + pid_c + 1)
b_ptr += chunk_seqlen_start * stride_b_seqlen + (
pid_h // nheads_ngroups_ratio) * stride_b_head pid_h // nheads_ngroups_ratio) * stride_b_head
x_ptr += pid_c * chunk_size * stride_x_seqlen + pid_h * stride_x_head x_ptr += chunk_seqlen_start * stride_x_seqlen + pid_h * stride_x_head
dt_ptr += pid_c * stride_dt_chunk + pid_h * stride_dt_head dt_ptr += pid_c * stride_dt_chunk + pid_h * stride_dt_head
dA_cumsum_ptr += pid_c * stride_dA_cs_chunk + pid_h * stride_dA_cs_head dA_cumsum_ptr += pid_c * stride_dA_cs_chunk + pid_h * stride_dA_cs_head
seq_idx_ptr += pid_c * chunk_size * stride_seq_idx_seqlen
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
offs_k = tl.arange(0, BLOCK_SIZE_K) offs_k = tl.arange(0, BLOCK_SIZE_K)
@ -243,10 +245,7 @@ def _chunk_state_fwd_kernel(
(chunk_size - 1) * stride_dA_cs_csize).to(tl.float32) (chunk_size - 1) * stride_dA_cs_csize).to(tl.float32)
dA_cumsum_ptrs = dA_cumsum_ptr + offs_k * stride_dA_cs_csize dA_cumsum_ptrs = dA_cumsum_ptr + offs_k * stride_dA_cs_csize
seq_idx_ptrs = seq_idx_ptr + offs_k * stride_seq_idx_seqlen chunk_size_limit = chunk_seqlen_end - chunk_seqlen_start
chunk_size_limit = min(chunk_size, seqlen - pid_c * chunk_size)
seq_idx_last = tl.load(seq_idx_ptr +
(chunk_size_limit - 1) * stride_seq_idx_seqlen)
acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, chunk_size_limit, BLOCK_SIZE_K): for k in range(0, chunk_size_limit, BLOCK_SIZE_K):
@ -261,15 +260,9 @@ def _chunk_state_fwd_kernel(
dA_cs_k = tl.load(dA_cumsum_ptrs, dA_cs_k = tl.load(dA_cumsum_ptrs,
mask=offs_k < chunk_size_limit - k, mask=offs_k < chunk_size_limit - k,
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
seq_idx_k = tl.load(seq_idx_ptrs,
mask=offs_k < chunk_size_limit - k,
other=-1)
dt_k = tl.load(dt_ptrs, mask=offs_k < chunk_size_limit - k, dt_k = tl.load(dt_ptrs, mask=offs_k < chunk_size_limit - k,
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
scale = tl.exp(dA_cs_last - dA_cs_k) * dt_k
scale = tl.where(seq_idx_k == seq_idx_last,
tl.exp(dA_cs_last - dA_cs_k) * dt_k, 0.0)
b *= scale[:, None] b *= scale[:, None]
b = b.to(x_ptr.dtype.element_ty) b = b.to(x_ptr.dtype.element_ty)
acc += tl.dot(x, b) acc += tl.dot(x, b)
@ -278,7 +271,6 @@ def _chunk_state_fwd_kernel(
b_ptrs += BLOCK_SIZE_K * stride_b_seqlen b_ptrs += BLOCK_SIZE_K * stride_b_seqlen
dt_ptrs += BLOCK_SIZE_K * stride_dt_csize dt_ptrs += BLOCK_SIZE_K * stride_dt_csize
dA_cumsum_ptrs += BLOCK_SIZE_K * stride_dA_cs_csize dA_cumsum_ptrs += BLOCK_SIZE_K * stride_dA_cs_csize
seq_idx_ptrs += BLOCK_SIZE_K * stride_seq_idx_seqlen
states = acc.to(states_ptr.dtype.element_ty) states = acc.to(states_ptr.dtype.element_ty)
@ -534,6 +526,7 @@ def _chunk_state_varlen_kernel(
def _chunk_cumsum_fwd(dt, def _chunk_cumsum_fwd(dt,
A, A,
chunk_size, chunk_size,
cu_chunk_seqlens,
dt_bias=None, dt_bias=None,
dt_softplus=False, dt_softplus=False,
dt_limit=(0.0, float("inf"))): dt_limit=(0.0, float("inf"))):
@ -541,7 +534,7 @@ def _chunk_cumsum_fwd(dt,
assert A.shape == (nheads, ) assert A.shape == (nheads, )
if dt_bias is not None: if dt_bias is not None:
assert dt_bias.shape == (nheads, ) assert dt_bias.shape == (nheads, )
nchunks = math.ceil(seqlen / chunk_size) nchunks = cu_chunk_seqlens.shape[0] - 1
dt_out = torch.empty(nheads, dt_out = torch.empty(nheads,
nchunks, nchunks,
chunk_size, chunk_size,
@ -561,6 +554,7 @@ def _chunk_cumsum_fwd(dt,
dt_bias_ptr=dt_bias, dt_bias_ptr=dt_bias,
dt_out_ptr=dt_out, dt_out_ptr=dt_out,
dA_cumsum_ptr=dA_cumsum, dA_cumsum_ptr=dA_cumsum,
cu_chunk_seqlens_ptr=cu_chunk_seqlens,
seqlen=seqlen, seqlen=seqlen,
nheads=nheads, nheads=nheads,
chunk_size=chunk_size, chunk_size=chunk_size,
@ -588,7 +582,7 @@ def _chunk_state_fwd(B,
x, x,
dt, dt,
dA_cumsum, dA_cumsum,
seq_idx=None, cu_chunk_seqlens,
states=None, states=None,
states_in_fp32=True): states_in_fp32=True):
seqlen, nheads, headdim = x.shape seqlen, nheads, headdim = x.shape
@ -599,9 +593,6 @@ def _chunk_state_fwd(B,
assert dt.shape == (nheads, nchunks, chunk_size) assert dt.shape == (nheads, nchunks, chunk_size)
assert dA_cumsum.shape == dt.shape assert dA_cumsum.shape == dt.shape
assert seq_idx is not None
assert seq_idx.shape == (seqlen, )
if states is not None: if states is not None:
assert states.shape == (nchunks, nheads, headdim, dstate) assert states.shape == (nchunks, nheads, headdim, dstate)
else: else:
@ -619,7 +610,7 @@ def _chunk_state_fwd(B,
states_ptr=states, states_ptr=states,
dt_ptr=dt, dt_ptr=dt,
dA_cumsum_ptr=dA_cumsum, dA_cumsum_ptr=dA_cumsum,
seq_idx_ptr=seq_idx, cu_chunk_seqlens_ptr=cu_chunk_seqlens,
hdim=headdim, hdim=headdim,
dstate=dstate, dstate=dstate,
chunk_size=chunk_size, chunk_size=chunk_size,
@ -641,7 +632,6 @@ def _chunk_state_fwd(B,
stride_dA_cs_head=dA_cumsum.stride(0), stride_dA_cs_head=dA_cumsum.stride(0),
stride_dA_cs_chunk=dA_cumsum.stride(1), stride_dA_cs_chunk=dA_cumsum.stride(1),
stride_dA_cs_csize=dA_cumsum.stride(2), stride_dA_cs_csize=dA_cumsum.stride(2),
stride_seq_idx_seqlen=seq_idx.stride(0),
) )
return states return states

56
vllm/model_executor/layers/mamba/ops/ssd_combined.py
View File

@ -14,8 +14,7 @@ from vllm.triton_utils import triton
from .ssd_bmm import _bmm_chunk_fwd from .ssd_bmm import _bmm_chunk_fwd
from .ssd_chunk_scan import _chunk_scan_fwd from .ssd_chunk_scan import _chunk_scan_fwd
from .ssd_chunk_state import (_chunk_cumsum_fwd, _chunk_state_fwd, from .ssd_chunk_state import _chunk_cumsum_fwd, _chunk_state_fwd
chunk_state_varlen)
from .ssd_state_passing import _state_passing_fwd from .ssd_state_passing import _state_passing_fwd
TRITON_22 = version.parse(triton.__version__) >= version.parse('2.2.0') TRITON_22 = version.parse(triton.__version__) >= version.parse('2.2.0')
@ -37,9 +36,9 @@ def _mamba_chunk_scan_combined_fwd(x,
dt_bias=None, dt_bias=None,
initial_states=None, initial_states=None,
seq_idx=None, seq_idx=None,
chunk_indices=None,
chunk_offsets=None,
cu_seqlens=None, cu_seqlens=None,
cu_chunk_seqlens=None,
last_chunk_indices=None,
dt_softplus=False, dt_softplus=False,
dt_limit=(0.0, float("inf")), dt_limit=(0.0, float("inf")),
state_dtype=None): state_dtype=None):
@ -56,7 +55,7 @@ def _mamba_chunk_scan_combined_fwd(x,
if D is not None: if D is not None:
assert D.shape == (nheads, headdim) or D.shape == (nheads, ) assert D.shape == (nheads, headdim) or D.shape == (nheads, )
if seq_idx is not None: if seq_idx is not None:
assert seq_idx.shape == (seqlen, ) assert seq_idx.shape == (cu_chunk_seqlens.shape[0] - 1, )
if B.stride(-1) != 1: if B.stride(-1) != 1:
B = B.contiguous() B = B.contiguous()
if C.stride(-1) != 1: if C.stride(-1) != 1:
@ -89,6 +88,7 @@ def _mamba_chunk_scan_combined_fwd(x,
dA_cumsum, dt = _chunk_cumsum_fwd(dt, dA_cumsum, dt = _chunk_cumsum_fwd(dt,
A, A,
chunk_size, chunk_size,
cu_chunk_seqlens,
dt_bias=dt_bias, dt_bias=dt_bias,
dt_softplus=dt_softplus, dt_softplus=dt_softplus,
dt_limit=dt_limit) dt_limit=dt_limit)
@ -99,36 +99,31 @@ def _mamba_chunk_scan_combined_fwd(x,
x, x,
dt, dt,
dA_cumsum, dA_cumsum,
seq_idx=seq_idx, cu_chunk_seqlens,
states_in_fp32=True) states_in_fp32=True)
# 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
# (middle term of factorization of off-diag blocks; A terms) # (middle term of factorization of off-diag blocks; A terms)
# - for handling chunked prefill, this requires i) initial_states # - for handling chunked prefill, this requires i) initial_states and
# ii) seq_idx iii) is_cont_batched and (iv) chunk_offsets to be all specified. # ii) seq_idx to be all specified.
# - When a new seq_idx is detected, we will stop passing the prev_state # - When a new seq_idx is detected, we will stop passing the prev_state
# and switch accordingly to the init_state corresponding to the new seq_idx. # and switch accordingly to the init_state corresponding to the new seq_idx.
# - We will also make sure that the dA_cumsum is taken only from the start of the
# sequence (hence we need the full dA_cumsum tensor and not just the values at chunk boundaries)
# - this will ensure that states will be updated with the rightmost flushed seq_idx
# of the previous chunk. This implies that the first chunk of states is either 0
# or equal to init_states of the first example.
states = _state_passing_fwd( states = _state_passing_fwd(
rearrange(states, "... p n -> ... (p n)"), rearrange(states, "... p n -> ... (p n)"),
dA_cumsum, # (nheads, nchunks, chunk_size) dA_cumsum, # (nheads, nchunks, chunk_size)
cu_chunk_seqlens,
initial_states=rearrange(initial_states, "... p n -> ... (p n)") initial_states=rearrange(initial_states, "... p n -> ... (p n)")
if initial_states is not None else if initial_states is not None else
None, # (batch, nheads, headdim*dstate) None, # (batch, nheads, headdim*dstate)
seq_idx=seq_idx, seq_idx=seq_idx,
out_dtype=state_dtype if state_dtype is not None else C.dtype, out_dtype=state_dtype if state_dtype is not None else C.dtype)
chunk_offsets=chunk_offsets)
states = rearrange(states, "... (p n) -> ... p n", n=dstate) states = rearrange(states, "... (p n) -> ... p n", n=dstate)
# 4. Compute batched matrix multiply for C_j^T B_i terms # 4. Compute batched matrix multiply for C_j^T B_i terms
CB = _bmm_chunk_fwd(C, CB = _bmm_chunk_fwd(C,
B, B,
chunk_size, chunk_size,
seq_idx=seq_idx, cu_chunk_seqlens,
output_dtype=torch.float32) output_dtype=torch.float32)
# 5. Scan and compute the diagonal blocks, taking into # 5. Scan and compute the diagonal blocks, taking into
@ -148,26 +143,15 @@ def _mamba_chunk_scan_combined_fwd(x,
dA_cumsum, dA_cumsum,
C, C,
states, states,
cu_chunk_seqlens,
out, # in-place update out, # in-place update
seq_idx, seq_idx,
D=D, D=D,
z=z, z=z,
chunk_indices=chunk_indices,
chunk_offsets=chunk_offsets,
initial_states=initial_states, initial_states=initial_states,
) )
varlen_states = chunk_state_varlen( return states[last_chunk_indices]
B,
x,
dt,
dA_cumsum,
cu_seqlens,
states,
initial_states=initial_states,
)
return varlen_states
def mamba_chunk_scan_combined_varlen( def mamba_chunk_scan_combined_varlen(
@ -178,14 +162,14 @@ def mamba_chunk_scan_combined_varlen(
C, C,
chunk_size, chunk_size,
cu_seqlens, cu_seqlens,
cu_chunk_seqlens,
last_chunk_indices,
seq_idx, seq_idx,
out, out,
D=None, D=None,
z=None, z=None,
dt_bias=None, dt_bias=None,
initial_states=None, initial_states=None,
chunk_indices=None,
chunk_offsets=None,
dt_softplus=False, dt_softplus=False,
dt_limit=(0.0, float("inf")), dt_limit=(0.0, float("inf")),
state_dtype=None, state_dtype=None,
@ -198,8 +182,10 @@ def mamba_chunk_scan_combined_varlen(
B: (seqlen, ngroups, dstate) B: (seqlen, ngroups, dstate)
C: (seqlen, ngroups, dstate) C: (seqlen, ngroups, dstate)
chunk_size: int chunk_size: int
seq_idx: (seqlen) cu_seqlens: (batch + 1,)
cu_seqlens: (batch + 1) cu_chunk_seqlens: (nchunks + 1,)
last_chunk_indices: (batch,)
seq_idx: (nchunks,)
out: (seqlen, nheads, headdim) preallocated output tensor out: (seqlen, nheads, headdim) preallocated output tensor
D: (nheads, headdim) or (nheads,) D: (nheads, headdim) or (nheads,)
z: (seqlen, nheads, headdim) z: (seqlen, nheads, headdim)
@ -228,9 +214,9 @@ def mamba_chunk_scan_combined_varlen(
dt_bias=dt_bias, dt_bias=dt_bias,
initial_states=initial_states, initial_states=initial_states,
seq_idx=seq_idx, seq_idx=seq_idx,
chunk_indices=chunk_indices,
chunk_offsets=chunk_offsets,
cu_seqlens=cu_seqlens, cu_seqlens=cu_seqlens,
cu_chunk_seqlens=cu_chunk_seqlens,
last_chunk_indices=last_chunk_indices,
dt_softplus=dt_softplus, dt_softplus=dt_softplus,
dt_limit=dt_limit, dt_limit=dt_limit,
state_dtype=state_dtype) state_dtype=state_dtype)

97
vllm/model_executor/layers/mamba/ops/ssd_state_passing.py
View File

@ -30,8 +30,7 @@ def _state_passing_fwd_kernel(
dA_cs_ptr, dA_cs_ptr,
initstates_ptr, initstates_ptr,
seq_idx_ptr, seq_idx_ptr,
chunk_offsets_ptr, cu_chunk_seqlens_ptr,
chunk_meta_num,
# Matrix dimensions # Matrix dimensions
dim: tl.constexpr, dim: tl.constexpr,
nchunks, nchunks,
@ -50,94 +49,52 @@ def _state_passing_fwd_kernel(
stride_initstates_batch: tl.int64, stride_initstates_batch: tl.int64,
stride_initstates_head: tl.int64, stride_initstates_head: tl.int64,
stride_initstates_dim: tl.constexpr, stride_initstates_dim: tl.constexpr,
stride_seq_idx_seqlen: tl.constexpr, stride_seq_idx_chunk: tl.constexpr,
# Meta-parameters # Meta-parameters
HAS_INITSTATES: tl.constexpr, HAS_INITSTATES: tl.constexpr,
BLOCK_SIZE: tl.constexpr, BLOCK_SIZE: tl.constexpr,
): ):
pid_h = tl.program_id(axis=1) pid_h = tl.program_id(axis=1)
pid_m = tl.program_id(axis=0) pid_m = tl.program_id(axis=0)
states_ptr += pid_h * stride_states_head states_ptr += pid_h * stride_states_head
dA_cs_ptr += pid_h * stride_dA_cs_head + (chunk_size - dA_cs_ptr += pid_h * stride_dA_cs_head + (chunk_size -
1) * stride_dA_cs_csize 1) * stride_dA_cs_csize
out_ptr += pid_h * stride_out_head out_ptr += pid_h * stride_out_head
if HAS_INITSTATES:
initstates_ptr += pid_h * stride_initstates_head
offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
states_ptrs = states_ptr + offs_m * stride_states_dim states_ptrs = states_ptr + offs_m * stride_states_dim
out_ptrs = out_ptr + offs_m * stride_out_dim out_ptrs = out_ptr + offs_m * stride_out_dim
# - states will be the past state of the sequence that continues on the current check if HAS_INITSTATES:
if not HAS_INITSTATES: initstates_ptrs = initstates_ptr \
states = tl.zeros((BLOCK_SIZE, ), dtype=tl.float32) + pid_h * stride_initstates_head \
else: + offs_m * stride_initstates_dim
initstates_ptr += offs_m * stride_initstates_dim
initstates_ptrs = initstates_ptr
# - for cont batches, for the first chunk mean it will be the first batch's
# init state
states = tl.load(initstates_ptrs, mask=offs_m < dim, states = tl.load(initstates_ptrs, mask=offs_m < dim,
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
else:
states = tl.zeros((BLOCK_SIZE, ), dtype=tl.float32)
tl.store(out_ptrs, states, mask=offs_m < dim) prev_seq_idx = 0
out_ptrs += stride_out_chunk for c in range(nchunks):
prev_seq_idx_chunk_end = 0
logical_chunk_idx = 0
for c in range(nchunks - 1):
new_states = tl.load(states_ptrs, mask=offs_m < dim, new_states = tl.load(states_ptrs, mask=offs_m < dim,
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
dA_cs = tl.load(dA_cs_ptr).to(tl.float32) dA_cs = tl.load(dA_cs_ptr).to(tl.float32)
scale_mask = True seq_idx = tl.load(seq_idx_ptr + c * stride_seq_idx_chunk)
# - the seq to pass forward is the one that is flushed to the right # we have started a new sequence
# boundary. if prev_seq_idx != seq_idx:
# - that is given by seq_idx_chunk_end below: the sequence index at the end of the chunk. if HAS_INITSTATES:
seq_idx_chunk_end = tl.load(seq_idx_ptr + initstates_ptrs = initstates_ptr + seq_idx * stride_initstates_batch \
(min((c + 1) * chunk_size, seqlen) - 1) * + pid_h * stride_initstates_head \
stride_seq_idx_seqlen) + offs_m * stride_initstates_dim
if HAS_INITSTATES:
if prev_seq_idx_chunk_end != seq_idx_chunk_end:
# this means in the current chunk the rightmost flushed seq
# has changed.
# - so we do not propagate the state from previous chunk
# - but rather we load that sequence's init state
initstates_ptrs = initstates_ptr + seq_idx_chunk_end * stride_initstates_batch
# - update state with seq_idx_new's init state
states = tl.load(initstates_ptrs, mask=offs_m < dim, states = tl.load(initstates_ptrs, mask=offs_m < dim,
other=0.0).to(tl.float32) other=0.0).to(tl.float32)
else:
states = tl.zeros((BLOCK_SIZE, ), dtype=tl.float32)
# - we need to consider the cumsum only of the last sequence in the chunk prev_seq_idx = seq_idx
# - find its starting position (given by c_off of the logical chunk index) states = tl.exp(dA_cs) * states + new_states
# - and subtract the cumsum just before that position from the total cumsum
# - first, update the logical chunk index (add the number of sequences in the current physical chunk):
# sequence index at the start of the current chunk
seq_idx_chunk_start = tl.load(seq_idx_ptr +
min(c * chunk_size, seqlen) *
stride_seq_idx_seqlen)
logical_chunk_idx += seq_idx_chunk_end - seq_idx_chunk_start
# - load the chunk offset:
c_off = tl.load(chunk_offsets_ptr + logical_chunk_idx,
mask=logical_chunk_idx < chunk_meta_num,
other=0)
# - if offset is 0, then the sequence starts at the beginning of the chunk, and we don't need to subtract anything
if c_off > 0:
# - dA_cs_ptr currently points to the cumsum at the end of the chunk - subtract the chunk size and add the offset
dA_cs_boundary = tl.load(
dA_cs_ptr - (chunk_size - 1) * stride_dA_cs_csize +
(c_off - 1) * stride_dA_cs_csize,
mask=(c_off - 1) > -1 and c_off < chunk_size,
other=0.0)
dA_cs -= dA_cs_boundary
# - increment logical chunk index for every physical chunk
logical_chunk_idx += 1
else:
scale_mask = seq_idx_chunk_end == prev_seq_idx_chunk_end
prev_seq_idx_chunk_end = seq_idx_chunk_end
scale = tl.where(scale_mask, tl.exp(dA_cs), 0.0)
states = scale * states + new_states
tl.store(out_ptrs, states, mask=offs_m < dim) tl.store(out_ptrs, states, mask=offs_m < dim)
states_ptrs += stride_states_chunk states_ptrs += stride_states_chunk
@ -148,8 +105,8 @@ def _state_passing_fwd_kernel(
def _state_passing_fwd( def _state_passing_fwd(
states, states,
dA_cumsum, dA_cumsum,
cu_chunk_seqlens,
seq_idx, seq_idx,
chunk_offsets,
initial_states=None, initial_states=None,
out_dtype=None, out_dtype=None,
): ):
@ -175,9 +132,7 @@ def _state_passing_fwd(
dA_cs_ptr=dA_cumsum, dA_cs_ptr=dA_cumsum,
initstates_ptr=initial_states, initstates_ptr=initial_states,
seq_idx_ptr=seq_idx, seq_idx_ptr=seq_idx,
chunk_offsets_ptr=chunk_offsets, cu_chunk_seqlens_ptr=cu_chunk_seqlens,
chunk_meta_num=len(chunk_offsets)
if chunk_offsets is not None else 0,
dim=dim, dim=dim,
nchunks=nchunks, nchunks=nchunks,
seqlen=seqlen if seq_idx is not None else 0, seqlen=seqlen if seq_idx is not None else 0,
@ -194,7 +149,7 @@ def _state_passing_fwd(
stride_initstates_batch=initial_states_strides[0], stride_initstates_batch=initial_states_strides[0],
stride_initstates_head=initial_states_strides[1], stride_initstates_head=initial_states_strides[1],
stride_initstates_dim=initial_states_strides[2], stride_initstates_dim=initial_states_strides[2],
stride_seq_idx_seqlen=seq_idx.stride(0), stride_seq_idx_chunk=seq_idx.stride(0),
HAS_INITSTATES=initial_states is not None, HAS_INITSTATES=initial_states is not None,
) )
return out return out

8
vllm/model_executor/models/plamo2.py
View File

@ -260,9 +260,9 @@ class Plamo2MambaMixer(MambaBase, CustomOp):
prep_initial_states = attn_metadata.prep_initial_states prep_initial_states = attn_metadata.prep_initial_states
chunk_size = attn_metadata.chunk_size chunk_size = attn_metadata.chunk_size
seq_idx_p = attn_metadata.seq_idx_p seq_idx_p = attn_metadata.seq_idx_p
chunk_indices_p = attn_metadata.chunk_indices_p
chunk_offsets_p = attn_metadata.chunk_offsets_p
query_start_loc_p = attn_metadata.query_start_loc_p query_start_loc_p = attn_metadata.query_start_loc_p
cu_chunk_seqlen_p = attn_metadata.cu_chunk_seqlen_p
last_chunk_indices_p = attn_metadata.last_chunk_indices_p
# 1. Gated MLP's linear projection # 1. Gated MLP's linear projection
projected_states = self.in_proj(hidden_states) projected_states = self.in_proj(hidden_states)
@ -368,9 +368,9 @@ class Plamo2MambaMixer(MambaBase, CustomOp):
self.num_heads // self.tp_size, self.head_dim), self.num_heads // self.tp_size, self.head_dim),
dt_bias=self.dt_bias, dt_bias=self.dt_bias,
seq_idx=seq_idx_p, seq_idx=seq_idx_p,
chunk_indices=chunk_indices_p,
chunk_offsets=chunk_offsets_p,
cu_seqlens=query_start_loc_p, cu_seqlens=query_start_loc_p,
cu_chunk_seqlens=cu_chunk_seqlen_p,
last_chunk_indices=last_chunk_indices_p,
initial_states=initial_states, initial_states=initial_states,
dt_softplus=True, dt_softplus=True,
dt_limit=(0.0, float("inf")), dt_limit=(0.0, float("inf")),

187
vllm/v1/attention/backends/mamba2_attn.py
View File

@ -1,6 +1,5 @@
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import math
from dataclasses import dataclass from dataclasses import dataclass
from typing import Optional from typing import Optional
@ -8,6 +7,7 @@ import torch
from vllm.attention.backends.abstract import AttentionBackend from vllm.attention.backends.abstract import AttentionBackend
from vllm.config import VllmConfig from vllm.config import VllmConfig
from vllm.utils import cdiv
from vllm.v1.attention.backends.mamba_attn import ( from vllm.v1.attention.backends.mamba_attn import (
BaseMambaAttentionMetadataBuilder) BaseMambaAttentionMetadataBuilder)
from vllm.v1.attention.backends.utils import (PAD_SLOT_ID, from vllm.v1.attention.backends.utils import (PAD_SLOT_ID,
@ -17,91 +17,6 @@ from vllm.v1.attention.backends.utils import (PAD_SLOT_ID,
from vllm.v1.kv_cache_interface import AttentionSpec from vllm.v1.kv_cache_interface import AttentionSpec
def _query_start_loc_to_chunk_indices_offsets(
query_start_loc: torch.Tensor, chunk_size: int,
total_seqlens: int) -> tuple[torch.Tensor, torch.Tensor]:
"""
Args:
query_start_loc (torch.Tensor): 1D tensor of cumulative sequence
lengths, shape (num_seqs + 1,).
The first element should be 0. Each entry represents the starting
index of a sequence in the flattened token array.
chunk_size (int): The size of each physical mamba chunk
(number of tokens per chunk).
total_seqlens (int): The total number of tokens in the batch.
Returns:
Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
- chunk_indices (torch.Tensor): 1D tensor of indices
indicating the physical chunk for each logical chunk.
- chunk_offsets (torch.Tensor): 1D tensor of offsets
indicating the starting index of each logical chunk within
its physical chunk.
This function computes the chunk indices and offsets for the given
query_start_loc and chunk_size. Both are tensors of integers with length N,
where N is the number of logical (pseudo) chunks.
A logical chunk is a sequence of tokens that are all part of the same
sequence and are all in the same physical mamba chunk.
In other words, a logical chunk changes every time we cross a sequence
boundary or a physical mamba chunk boundary.
Logical chunks are needed to handle batched requests with initial states
(see _state_passing_fwd and _chunk_scan_fwd).
The chunk_indices tensor contains the index of the physical chunk for each
logical chunk.
The chunk_offsets tensor contains the offset (AKA starting index) of the
logical chunk in the physical chunk.
Example:
query_start_loc = [0, 5, 10]
chunk_size = 8
total_seqlens = 10
-> chunk_indices = [0, 0, 1]
-> chunk_offsets = [0, 5, 0]
In this example, we have 2 sequences, each with 5 tokens. The physical
chunk size is 8 tokens.
We have three logical chunks:
- the first logical chunk starts at token 0 in the first physical chunk
and contains all 5 tokens from the first sequence
- the second logical chunk starts at token 5 in the first physical chunk
and contains first 3 tokens from the second sequence
- the third logical chunk starts at token 0 in the second physical chunk
and contains the remaining 2 tokens from the second sequence
"""
cu_seqlens = query_start_loc[1:] # remove prepended 0
# outputs will have length expansion of chunks that do not divide
# chunk_size
N = math.ceil(total_seqlens / chunk_size) + (cu_seqlens[:-1] % chunk_size
> 0).sum()
chunk_indices = torch.arange(N,
dtype=torch.int,
device=query_start_loc.device)
chunk_offsets = torch.zeros((N, ),
dtype=torch.int,
device=query_start_loc.device)
p = 0 # num of insertions
for s, e in zip(cu_seqlens[:-1], cu_seqlens[1:]):
# if does not divide chunk_size, then there is one chunk insertion
p += (s % chunk_size > 0)
# get the dimensions
# - the + 1 for _e is to shift the boundary by one chunk
# - this shifting is not needed if chunk_size divides e
_s, _e = s // chunk_size + p, e // chunk_size + p + (e % chunk_size
> 0)
# adjust indices and offsets
chunk_indices[_s:_e] -= p
chunk_offsets[_s] = s % chunk_size
return chunk_indices, chunk_offsets
class Mamba2AttentionBackend(AttentionBackend): class Mamba2AttentionBackend(AttentionBackend):
@staticmethod @staticmethod
@ -125,8 +40,16 @@ class Mamba2AttentionMetadata:
# the batch has no prefill request. # the batch has no prefill request.
has_initial_states_p: Optional[torch.Tensor] has_initial_states_p: Optional[torch.Tensor]
seq_idx_p: Optional[torch.Tensor] seq_idx_p: Optional[torch.Tensor]
chunk_indices_p: Optional[torch.Tensor]
chunk_offsets_p: Optional[torch.Tensor] # cu_chunk_seqlen_p is a tensor of shape (nchunks+1,) that contains, for
# each chunk, its offests into the varlen sequence dimension. It is defined
# such that the i-th chunk contains tokens from cu_chunk_seqlen_p[i] to
# cu_chunk_seqlen_p[i+1].
cu_chunk_seqlen_p: Optional[torch.Tensor]
# last_chunk_indices_p is a tensor of shape (batch,) that contains the
# index of the last chunk for every sequence in the (prefill) batch.
last_chunk_indices_p: Optional[torch.Tensor]
state_indices_tensor: torch.Tensor # shape: [batch,] state_indices_tensor: torch.Tensor # shape: [batch,]
@ -151,13 +74,14 @@ class Mamba2AttentionMetadataBuilder(
common_attn_metadata: CommonAttentionMetadata, common_attn_metadata: CommonAttentionMetadata,
fast_build: bool = False) -> Mamba2AttentionMetadata: fast_build: bool = False) -> Mamba2AttentionMetadata:
num_reqs = common_attn_metadata.num_reqs num_reqs = common_attn_metadata.num_reqs
query_start_loc_p = None
seq_lens = common_attn_metadata.seq_lens seq_lens = common_attn_metadata.seq_lens
query_start_loc_p = None
seq_idx_p = None seq_idx_p = None
chunk_indices_p, chunk_offsets_p = None, None cu_chunk_seqlen_p = None
last_chunk_indices_p = None
# Need flags to indicate if there are initial states # Need flags to indicate if there are initial states
# currently we really only support the FlashAttention backend
has_initial_states_p = None has_initial_states_p = None
prep_initial_states = False prep_initial_states = False
@ -171,7 +95,7 @@ class Mamba2AttentionMetadataBuilder(
common_attn_metadata, common_attn_metadata,
decode_threshold=self.reorder_batch_threshold)) decode_threshold=self.reorder_batch_threshold))
# Compute seq_idx, chunk_indices and chunk_offsets for prefill only # Compute seq_idx for prefill only
if num_prefills > 0: if num_prefills > 0:
#[batch,] #[batch,]
has_initial_states_cpu = ( has_initial_states_cpu = (
@ -184,21 +108,68 @@ class Mamba2AttentionMetadataBuilder(
query_start_loc_p = common_attn_metadata.query_start_loc[ query_start_loc_p = common_attn_metadata.query_start_loc[
-num_prefills - 1:] - num_decode_tokens -num_prefills - 1:] - num_decode_tokens
seq_idx_p = torch.repeat_interleave(torch.arange( num_computed_tokens_p = \
num_prefills, common_attn_metadata.num_computed_tokens_cpu[
dtype=torch.int32, num_reqs - num_prefills:num_reqs]
device=query_start_loc_p.device), query_start_loc_p_cpu = common_attn_metadata.query_start_loc_cpu[
query_start_loc_p.diff(), -num_prefills - 1:] - num_decode_tokens
output_size=num_prefill_tokens)
# We compute metadata for chunked prefill once at the top level # The code below carefully constructs the chunks such that:
# model forward and reuse them in mamba layers. If not needed, # 1. Chunks contain tokens from a *single* sequence only.
# they will be ignored inside mamba kernels. # 2. For every sequence, we are guaranteed that we can
if prep_initial_states: # retrieve the mamba state *every* chunk_size tokens.
chunk_indices_p, chunk_offsets_p = ( # Constraint (1) dramatically simplifies the mamba2 kernels.
_query_start_loc_to_chunk_indices_offsets( # Constraint (2) dramatically simplifies the implementation
query_start_loc_p, self.chunk_size, # of prefix caching for mamba2 (wip). We need to take care
num_prefill_tokens)) # of the interaction with chunked prefill in order to
# satisfy constraint (2).
# TODO (tdoublep): This code could probably be optimized.
cu_chunk_seqlen = []
seq_idx = []
last_chunk_indices = []
seqlen_pos = 0
for req_idx in range(num_prefills):
this_num_computed = num_computed_tokens_p[req_idx].item()
this_new_tokens = query_start_loc_p_cpu[req_idx + 1].item(
) - query_start_loc_p_cpu[req_idx].item()
# if computed tokens are not chunk-aligned, use the first
# chunk to finish it off
if this_num_computed % self.chunk_size != 0:
seq_idx.append(req_idx)
cu_chunk_seqlen.append(seqlen_pos)
# how many tokens to finish the chunk?
chunk_len = cdiv(this_num_computed, self.chunk_size
) * self.chunk_size - this_num_computed
# we can only use at most this_new_tokens
chunk_len = min(chunk_len, this_new_tokens)
seqlen_pos += chunk_len
this_new_tokens -= chunk_len
n_chunks = cdiv(this_new_tokens, self.chunk_size)
for chunk in range(n_chunks):
seq_idx.append(req_idx)
cu_chunk_seqlen.append(seqlen_pos)
chunk_len = min(self.chunk_size, this_new_tokens)
seqlen_pos += chunk_len
this_new_tokens -= chunk_len
assert this_new_tokens == 0
last_chunk_indices.append(len(cu_chunk_seqlen) - 1)
cu_chunk_seqlen.append(seqlen_pos)
seq_idx_p = torch.as_tensor(seq_idx,
device=query_start_loc_p.device,
dtype=torch.int32)
cu_chunk_seqlen_p = torch.as_tensor(
cu_chunk_seqlen,
device=query_start_loc_p.device,
dtype=torch.int32)
last_chunk_indices_p = torch.as_tensor(
last_chunk_indices,
device=query_start_loc_p.device,
dtype=torch.int32)
nums_dict, batch_ptr, token_chunk_offset_ptr = \ nums_dict, batch_ptr, token_chunk_offset_ptr = \
compute_causal_conv1d_metadata(query_start_loc_p) compute_causal_conv1d_metadata(query_start_loc_p)
@ -222,9 +193,9 @@ class Mamba2AttentionMetadataBuilder(
chunk_size=self.chunk_size, chunk_size=self.chunk_size,
has_initial_states_p=has_initial_states_p, has_initial_states_p=has_initial_states_p,
seq_idx_p=seq_idx_p, seq_idx_p=seq_idx_p,
chunk_indices_p=chunk_indices_p,
chunk_offsets_p=chunk_offsets_p,
state_indices_tensor=state_indices_tensor, state_indices_tensor=state_indices_tensor,
cu_chunk_seqlen_p=cu_chunk_seqlen_p,
last_chunk_indices_p=last_chunk_indices_p,
nums_dict=nums_dict, nums_dict=nums_dict,
batch_ptr=batch_ptr, batch_ptr=batch_ptr,
token_chunk_offset_ptr=token_chunk_offset_ptr, token_chunk_offset_ptr=token_chunk_offset_ptr,