[Model] add GLM-5.3-Flash support (#53906)

This commit is contained in:
Jiangyun Zhu
2026-09-03 09:40:35 -07:00
committed by GitHub
parent da8ec28268
commit 98ed0856f3
94 changed files with 17927 additions and 266 deletions
+2
View File
@@ -0,0 +1,2 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+2
View File
@@ -0,0 +1,2 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
@@ -0,0 +1,890 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""kpool (key-pooling) Triton kernels for the sparse-attention indexer.
The cache stores POOLS (1 entry per ``pool_size`` consecutive tokens) rather
than individual tokens. ``compress_ratio == pool_size`` on the kv_cache_spec
makes the metadata builder emit pool-granular slot_mapping / seq_lens /
cu_seq_lens / page_table for free; this file supplies the compress-write
kernel (replacing ``indexer_k_quant_and_cache``) and the pool-level topk
helpers (select pools -> expand to tokens -> append tail).
"""
from __future__ import annotations
import torch
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
# The GLM-5.3-Flash indexer head dimension is fixed at 128.
INDEX_HEAD_DIM = 128
FP8_DTYPE = current_platform.fp8_dtype()
FP8_MAX = torch.finfo(FP8_DTYPE).max
@triton.jit
def _cache_k_offset(
token_offset,
dim_offset,
head_dim: tl.constexpr,
preshuffle: tl.constexpr,
):
if preshuffle:
return (
(token_offset // 16) * 16 * head_dim
+ (dim_offset // 16) * 16 * 16
+ (token_offset % 16) * 16
+ dim_offset % 16
)
return token_offset * head_dim + dim_offset
# Hadamard-128 rotation
@triton.jit
def _hadamard128_stage(x, GROUPS: tl.constexpr, STRIDE: tl.constexpr):
x3 = tl.reshape(x, (GROUPS, 2, STRIDE))
x3 = tl.trans(x3, 0, 2, 1)
a, b = tl.split(x3)
x3 = tl.join(a + b, a - b)
x3 = tl.trans(x3, 0, 2, 1)
return tl.reshape(x3, (128,))
@triton.jit
def _hadamard128(x):
x = _hadamard128_stage(x, 64, 1)
x = _hadamard128_stage(x, 32, 2)
x = _hadamard128_stage(x, 16, 4)
x = _hadamard128_stage(x, 8, 8)
x = _hadamard128_stage(x, 4, 16)
x = _hadamard128_stage(x, 2, 32)
x = _hadamard128_stage(x, 1, 64)
return x * 0.08838834764831845 # 1/sqrt(128)
# Map pool ids to physical cache slots.
def compute_pooled_write_locs(
page_table_64: torch.Tensor,
pool_ids: torch.Tensor,
pool_size: int,
) -> torch.Tensor:
"""Map logical pooled-K ids to physical flat cache slots.
``pool_size`` consecutive tokens share one pool slot that lives at the
*first* token page of each page-group. ``page_table_64`` maps token pages
to physical block ids; we gather the block id of each pool's page-group
and add the in-block pool offset.
"""
assert page_table_64.ndim == 1
pool_ids = pool_ids.to(torch.int64)
block_size = 64
pool_page_group = torch.div(pool_ids, block_size, rounding_mode="floor")
token_page_row = pool_page_group * pool_size
packed_page = page_table_64.index_select(0, token_page_row.to(torch.int64))
return packed_page.to(torch.int64) * block_size + torch.remainder(
pool_ids, block_size
)
def build_pooled_page_table(
page_table: torch.Tensor,
pool_size: int,
) -> torch.Tensor:
"""Build a pool-granular page table by taking every ``pool_size``-th
token-page column (one pool maps to ``pool_size`` token pages).
Uses gather (not strided slicing) so the result is always a fresh
row-major tensor — some downstream kernels require stride(-1) == 1.
"""
block_size = page_table.shape[-1]
assert block_size % pool_size == 0, (
f"pool_size ({pool_size}) must divide page columns ({block_size})"
)
idx = torch.arange(0, block_size, pool_size, device=page_table.device)
return page_table[..., idx].contiguous()
# Fused pool compression and cache write.
@triton.jit
def _kpool_softmax_rotate_write_cache_kernel(
buf_fp8_ptr,
buf_fp32_ptr,
slot_k_ptr,
slot_score_ptr,
ape_ptr,
loc_ptr,
write_mask_ptr,
compressed_k_ptr,
compressed_scale_ptr,
slot_k_stride_0,
slot_k_stride_1,
slot_score_stride_0,
slot_score_stride_1,
ape_stride_0,
PAGE_SIZE: tl.constexpr,
BUF_NUMEL_PER_PAGE: tl.constexpr,
POOL_SIZE: tl.constexpr,
HEAD_DIM: tl.constexpr,
S_OFFSET_NBYTES_IN_PAGE: tl.constexpr,
FP8_MAX: tl.constexpr,
PRESHUFFLE: tl.constexpr,
ROUND_SCALE: tl.constexpr,
HAS_WRITE_MASK: tl.constexpr,
RETURN_COMPRESSED: tl.constexpr,
WRITE_CACHE: tl.constexpr,
BLOCK_D: tl.constexpr,
):
"""One program per pool. softmax(slot_score+ape)-weighted sum of slot_k ->
Hadamard-128 -> per-vector fp8 absmax quant -> write to cache at ``loc``."""
row = tl.program_id(0)
do_write = True
if HAS_WRITE_MASK:
do_write = tl.load(write_mask_ptr + row)
offs = tl.arange(0, BLOCK_D)
mask = (offs < HEAD_DIM) & do_write
# --- Pass 1: per-dim max over the pool (softmax numerical stability) ---
max_score = tl.full((BLOCK_D,), -float("inf"), tl.float32)
for slot in tl.static_range(0, POOL_SIZE):
score = tl.load(
slot_score_ptr
+ row * slot_score_stride_0
+ slot * slot_score_stride_1
+ offs,
mask=mask,
other=0.0,
).to(tl.float32)
score += tl.load(ape_ptr + slot * ape_stride_0 + offs, mask=mask, other=0.0).to(
tl.float32
)
max_score = tl.maximum(max_score, score)
# --- Pass 2: softmax-weighted sum of K ---
acc = tl.full((BLOCK_D,), 0.0, tl.float32)
denom = tl.full((BLOCK_D,), 0.0, tl.float32)
for slot in tl.static_range(0, POOL_SIZE):
score = tl.load(
slot_score_ptr
+ row * slot_score_stride_0
+ slot * slot_score_stride_1
+ offs,
mask=mask,
other=0.0,
).to(tl.float32)
score += tl.load(ape_ptr + slot * ape_stride_0 + offs, mask=mask, other=0.0).to(
tl.float32
)
prob = tl.exp(score - max_score)
denom += prob
k = tl.load(
slot_k_ptr + row * slot_k_stride_0 + slot * slot_k_stride_1 + offs,
mask=mask,
other=0.0,
).to(tl.float32)
acc += k * prob
x = acc / denom
x = tl.where(do_write, x, 0.0).to(tl.bfloat16).to(tl.float32)
# Match the unfused pooled-K path's bf16 precision before quantization.
x = _hadamard128(x).to(tl.bfloat16).to(tl.float32)
# --- per-vector absmax fp8 quant ---
fp8_max_inv = 1.0 / FP8_MAX
absmax = tl.max(tl.abs(x), axis=0)
absmax = tl.maximum(absmax, 1e-4)
if ROUND_SCALE:
scale = tl.exp2(tl.ceil(tl.log2(absmax * fp8_max_inv)))
else:
scale = absmax * fp8_max_inv
quantized = x / scale
quantized = tl.minimum(tl.maximum(quantized, -FP8_MAX), FP8_MAX)
if WRITE_CACHE:
loc = tl.load(loc_ptr + row, mask=do_write, other=0)
loc_page_index = loc // PAGE_SIZE
loc_token_offset_in_page = loc % PAGE_SIZE
out_k_offsets = loc_page_index * BUF_NUMEL_PER_PAGE + _cache_k_offset(
loc_token_offset_in_page,
offs,
HEAD_DIM,
PRESHUFFLE,
)
out_s_offset = (
loc_page_index * BUF_NUMEL_PER_PAGE // 4
+ S_OFFSET_NBYTES_IN_PAGE // 4
+ loc_token_offset_in_page
)
tl.store(buf_fp8_ptr + out_k_offsets, quantized, mask=mask)
tl.store(buf_fp32_ptr + out_s_offset, scale, mask=do_write)
if RETURN_COMPRESSED:
tl.store(
compressed_k_ptr + row * HEAD_DIM + offs,
quantized,
mask=offs < HEAD_DIM,
)
tl.store(compressed_scale_ptr + row, scale)
def kpool_compress_and_write_cache(
kv_cache: torch.Tensor,
slot_k: torch.Tensor,
slot_score: torch.Tensor,
ape: torch.Tensor,
loc: torch.Tensor,
pool_size: int,
head_dim: int = INDEX_HEAD_DIM,
write_mask: torch.Tensor | None = None,
round_scale: bool = True,
return_compressed: bool = False,
write_cache: bool = True,
):
"""Compress ``pool_size`` tokens into one fp8 K and write at ``loc``.
Args:
kv_cache: indexer K cache ``[num_blocks, block_size, head_dim+4]`` uint8.
slot_k: ``[n_pools, pool_size, head_dim]`` bf16 — raw per-token K.
slot_score: ``[n_pools, pool_size, head_dim]`` — per-token gate score.
ape: ``[pool_size, head_dim]`` fp32 — per-slot position bias.
loc: ``[n_pools]`` int64 — flat physical slot per pool.
"""
assert slot_k.ndim == 3
assert slot_score.shape == slot_k.shape
assert ape.shape == slot_k.shape[1:]
assert slot_k.shape[2] == head_dim
assert slot_k.dtype == torch.bfloat16
assert ape.dtype == torch.float32
assert kv_cache.dtype == torch.uint8
assert loc.dtype == torch.int64
assert write_cache or return_compressed
page_size = kv_cache.shape[1]
buf = kv_cache
slot_k = slot_k.contiguous()
slot_score = slot_score.contiguous()
ape = ape.contiguous()
loc = loc.contiguous()
if write_mask is None:
write_mask = torch.empty((1,), dtype=torch.bool, device=slot_k.device)
has_write_mask = False
else:
assert write_mask.shape == (slot_k.shape[0],)
write_mask = write_mask.contiguous()
has_write_mask = True
assert not return_compressed
if slot_k.shape[0] == 0:
if return_compressed:
return (
torch.empty(
(0, head_dim),
dtype=FP8_DTYPE,
device=slot_k.device,
),
torch.empty((0,), dtype=torch.float32, device=slot_k.device),
)
return None
buf_fp8 = buf.view(FP8_DTYPE)
buf_fp32 = buf.view(torch.float32)
# bytes per page (last dim of kv_cache) viewed as uint8
buf_numel_per_page = buf.stride(0)
s_offset_nbytes_in_page = page_size * head_dim
if return_compressed:
compressed_k = torch.empty(
(slot_k.shape[0], head_dim),
dtype=FP8_DTYPE,
device=slot_k.device,
)
compressed_scale = torch.empty(
(slot_k.shape[0],), dtype=torch.float32, device=slot_k.device
)
else:
compressed_k = buf_fp8
compressed_scale = buf_fp32
if page_size > 1:
assert page_size % 16 == 0, "ROCm preshuffle requires 16-token tiles"
_kpool_softmax_rotate_write_cache_kernel[(slot_k.shape[0],)](
buf_fp8,
buf_fp32,
slot_k,
slot_score,
ape,
loc,
write_mask,
compressed_k,
compressed_scale,
slot_k.stride(0),
slot_k.stride(1),
slot_score.stride(0),
slot_score.stride(1),
ape.stride(0),
PAGE_SIZE=page_size,
BUF_NUMEL_PER_PAGE=buf_numel_per_page,
POOL_SIZE=slot_k.shape[1],
HEAD_DIM=head_dim,
S_OFFSET_NBYTES_IN_PAGE=s_offset_nbytes_in_page,
FP8_MAX=FP8_MAX,
PRESHUFFLE=page_size > 1,
ROUND_SCALE=round_scale,
HAS_WRITE_MASK=has_write_mask,
RETURN_COMPRESSED=return_compressed,
WRITE_CACHE=write_cache,
BLOCK_D=triton.next_power_of_2(head_dim),
)
if return_compressed:
return compressed_k, compressed_scale
return None
# Seed each request's incomplete pool into its paged tail during prefill.
@triton.jit
def _kpool_tail_seed_kernel(
key_ptr,
score_ptr,
tslot_ptr,
tail_ptr,
n_tokens,
TAIL_BLOCK_ELEMS: tl.constexpr,
KPOOL_HEAD: tl.constexpr,
HEAD_DIM: tl.constexpr,
KPOOL: tl.constexpr,
BLOCK_D: tl.constexpr,
):
"""Copy token ``i``'s raw K + gate into its request's tail block.
Token ``i`` is among its request's last KPOOL tokens iff the token KPOOL
ahead belongs to a different tail block (or is past the batch / padding,
slot < 0). ``tslot = block * KPOOL + pos % KPOOL``; the destination is
``tail[block, {0:K, 1:score}, pos % KPOOL, :]``.
"""
i = tl.program_id(0)
t = tl.load(tslot_ptr + i).to(tl.int64)
if t < 0:
return
blk = t // KPOOL # t >= 0 here, so trunc == floor
ahead = tl.load(tslot_ptr + i + KPOOL, mask=i + KPOOL < n_tokens, other=-1).to(
tl.int64
)
# Match the torch semantics exactly: a negative ahead slot floors to a
# block id that differs from every real block -> token is in the tail.
# Only divide non-negative slots (Triton int div truncates, torch floors).
if ahead >= 0 and ahead // KPOOL == blk:
return
offs = tl.arange(0, BLOCK_D)
m = offs < HEAD_DIM
block_base = blk * TAIL_BLOCK_ELEMS
base = block_base + (t % KPOOL) * HEAD_DIM
k = tl.load(key_ptr + i * HEAD_DIM + offs, mask=m)
s = tl.load(score_ptr + i * HEAD_DIM + offs, mask=m)
tl.store(tail_ptr + base + offs, k, mask=m)
tl.store(
tail_ptr + block_base + KPOOL_HEAD + (t % KPOOL) * HEAD_DIM + offs, s, mask=m
)
def kpool_seed_tail_cache(
tail_kv_cache: torch.Tensor,
key: torch.Tensor,
gate_score: torch.Tensor,
tslot: torch.Tensor,
kpool: int,
head_dim: int = INDEX_HEAD_DIM,
) -> None:
"""Seed the paged tail cache from a prefill batch (see the kernel)."""
assert tail_kv_cache.dtype == torch.bfloat16
assert key.dtype == torch.bfloat16
n = tslot.shape[0]
if n == 0:
return
_kpool_tail_seed_kernel[(n,)](
key,
gate_score,
tslot,
tail_kv_cache,
n,
TAIL_BLOCK_ELEMS=tail_kv_cache.stride(0),
KPOOL_HEAD=tail_kv_cache.stride(1),
HEAD_DIM=head_dim,
KPOOL=kpool,
BLOCK_D=triton.next_power_of_2(head_dim),
)
# Update each request's tail during decode and write completed pools.
@triton.jit
def _kpool_decode_update_batched_kernel(
buf_fp8_ptr,
buf_fp32_ptr,
tail_kv_ptr,
tail_slot_mapping_ptr, # [B, NEXT_N] int32
key_ptr, # [B, NEXT_N, HEAD_DIM] bf16
key_stride_b,
key_stride_t,
slot_score_ptr, # [B, NEXT_N, HEAD_DIM] bf16
ss_stride_b,
ss_stride_t,
ape_ptr,
ape_stride_0,
slot_mapping_ptr, # [B, NEXT_N] int32
positions_ptr, # [B, NEXT_N] int32
NEXT_N, # runtime token count per request (no .item() needed)
PAGE_SIZE: tl.constexpr,
BUF_NUMEL_PER_PAGE: tl.constexpr,
POOL_SIZE: tl.constexpr,
TAIL_BLOCK_ELEMS: tl.constexpr,
KPOOL_HEAD: tl.constexpr,
HEAD_DIM: tl.constexpr,
S_OFFSET_NBYTES_IN_PAGE: tl.constexpr,
FP8_MAX: tl.constexpr,
PRESHUFFLE: tl.constexpr,
ROUND_SCALE: tl.constexpr,
BLOCK_D: tl.constexpr,
):
"""One program per request; iterates its NEXT_N verify tokens in order.
Replaces the caller's per-token sequential launch loop. The intra-request
iteration MUST stay in position order: a pool-completion at token t* reads
the tail-ring slots that tokens t < t* (same request) just stashed in this
same invocation. ``tl.range`` iterates sequentially within the program, so
those stashes are visible to the later completion read. Cross-request
programs are independent (distinct tail blocks). With NEXT_N < POOL_SIZE
(the spec-verify case: NEXT_N ~= num_spec+1, POOL_SIZE=16) at most one
completion can occur per request per call, but the ordered loop is correct
for any NEXT_N.
"""
req = tl.program_id(0)
offs = tl.arange(0, BLOCK_D)
dim_mask = offs < HEAD_DIM
for t in tl.range(0, NEXT_N):
idx = req * NEXT_N + t
cache_loc = tl.load(slot_mapping_ptr + idx)
pos = tl.load(positions_ptr + idx)
safe_pos = tl.maximum(pos, 0)
pos_valid = (cache_loc >= 0) & (pos >= 0)
slot = safe_pos % POOL_SIZE
phys_slot = safe_pos % POOL_SIZE
# Derive the tail block from THIS token's tail_slot (the request's block
# is constant across a pool, but a padded / invalid entry carries a
# negative sentinel -- reading it from token 0 would poison every
# token's base address). Clamp so an invalid entry can never form an
# out-of-bounds base; the accesses below are gated on pos_valid anyway.
tail_slot = tl.load(tail_slot_mapping_ptr + idx)
block = tl.maximum(tail_slot, 0).to(tl.int64) // POOL_SIZE
block_base = block * TAIL_BLOCK_ELEMS
# The tail-ring stash must run for EVERY real token, so it is gated on
# the token-granular tail slot -- not on `pos_valid`, which keys off the
# POOL-granular `slot_mapping` and is therefore only true on the pool's
# last token. Gating the stash on pos_valid dropped every intra-pool
# token, so a decode-built pool compressed 3 stale ring entries (the
# prefill-seeded prompt tail, frozen forever) plus the current token.
stash_valid = (pos >= 0) & (tail_slot >= 0)
key = tl.load(
key_ptr + req * key_stride_b + t * key_stride_t + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
score_current = tl.load(
slot_score_ptr + req * ss_stride_b + t * ss_stride_t + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
if pos_valid & (slot == POOL_SIZE - 1):
pool_logical_start = safe_pos - slot
max_score = tl.full((BLOCK_D,), -float("inf"), tl.float32)
for pool_slot in tl.static_range(0, POOL_SIZE):
is_current = pool_slot == slot
phys = (pool_logical_start + pool_slot) % POOL_SIZE
score_buf = tl.load(
tail_kv_ptr + block_base + KPOOL_HEAD + phys * HEAD_DIM + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
score = tl.where(is_current, score_current, score_buf)
score += tl.load(
ape_ptr + pool_slot * ape_stride_0 + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
max_score = tl.maximum(max_score, score)
acc = tl.full((BLOCK_D,), 0.0, tl.float32)
denom = tl.full((BLOCK_D,), 0.0, tl.float32)
for pool_slot in tl.static_range(0, POOL_SIZE):
is_current = pool_slot == slot
phys = (pool_logical_start + pool_slot) % POOL_SIZE
score_buf = tl.load(
tail_kv_ptr + block_base + KPOOL_HEAD + phys * HEAD_DIM + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
score = tl.where(is_current, score_current, score_buf)
score += tl.load(
ape_ptr + pool_slot * ape_stride_0 + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
prob = tl.exp(score - max_score)
denom += prob
k_buf = tl.load(
tail_kv_ptr + block_base + phys * HEAD_DIM + offs,
mask=dim_mask,
other=0.0,
).to(tl.float32)
k = tl.where(is_current, key, k_buf)
acc += k * prob
x = (acc / denom).to(tl.bfloat16).to(tl.float32)
x = _hadamard128(x).to(tl.bfloat16).to(tl.float32)
fp8_max_inv = 1.0 / FP8_MAX
absmax = tl.maximum(tl.max(tl.abs(x), axis=0), 1e-4)
if ROUND_SCALE:
scale = tl.exp2(tl.ceil(tl.log2(absmax * fp8_max_inv)))
else:
scale = absmax * fp8_max_inv
quantized = tl.minimum(tl.maximum(x / scale, -FP8_MAX), FP8_MAX)
loc = cache_loc.to(tl.int64)
loc_page_index = loc // PAGE_SIZE
loc_token_offset_in_page = loc % PAGE_SIZE
out_k_offsets = loc_page_index * BUF_NUMEL_PER_PAGE + _cache_k_offset(
loc_token_offset_in_page,
offs,
HEAD_DIM,
PRESHUFFLE,
)
out_s_offset = (
loc_page_index * BUF_NUMEL_PER_PAGE // 4
+ S_OFFSET_NBYTES_IN_PAGE // 4
+ loc_token_offset_in_page
)
tl.store(buf_fp8_ptr + out_k_offsets, quantized, mask=dim_mask)
tl.store(buf_fp32_ptr + out_s_offset, scale)
# Stash the current token AFTER any completion read so the completion
# uses prior stashes (and the current token's own key/score via
# is_current), then leaves this token for future pools. Order matches
# the per-token kernel: completion read first, stash second.
update_mask = dim_mask & stash_valid
tl.store(
tail_kv_ptr + block_base + phys_slot * HEAD_DIM + offs,
key,
mask=update_mask,
)
tl.store(
tail_kv_ptr + block_base + KPOOL_HEAD + phys_slot * HEAD_DIM + offs,
score_current,
mask=update_mask,
)
def kpool_decode_update_and_maybe_write_cache_batched(
kv_cache: torch.Tensor,
tail_kv_cache: torch.Tensor,
tail_slot_mapping: torch.Tensor,
key: torch.Tensor,
slot_score: torch.Tensor,
ape: torch.Tensor,
slot_mapping: torch.Tensor,
positions: torch.Tensor,
pool_size: int,
head_dim: int = INDEX_HEAD_DIM,
round_scale: bool = True,
) -> None:
"""Batched decode-step kpool update for spec verify (``next_n > 1``).
One launch replaces the caller's per-token loop. Inputs are grouped per
request: ``[num_requests, next_n, ...]``. Each program handles one
request's ``next_n`` tokens in position order (see the kernel docstring for
why ordering is required for pool-completion correctness).
Plain decode (``next_n == 1``) is handled here too — the kernel collapses
to a single-iteration loop.
Args:
kv_cache: indexer K cache ``[num_blocks, block_size, head_dim+4]`` uint8.
tail_kv_cache: paged tail cache ``[num_blocks, 2, pool_size, head_dim]``
bf16 (K at half 0, gate score at half 1).
tail_slot_mapping: ``[num_requests, next_n]`` int32.
key / slot_score: ``[num_requests, next_n, head_dim]`` bf16.
ape: ``[pool_size, head_dim]`` fp32.
slot_mapping / positions: ``[num_requests, next_n]`` int32.
"""
num_requests, next_n = key.shape[0], key.shape[1]
if num_requests == 0 or next_n == 0:
return
assert tail_kv_cache.ndim == 4
assert tail_kv_cache.shape[1] == 2
assert tail_kv_cache.shape[2] == pool_size
assert tail_kv_cache.shape[3] == head_dim
assert tail_kv_cache.dtype == torch.bfloat16
assert key.ndim == 3 and key.shape[2] == head_dim
assert slot_score.shape == key.shape
assert ape.shape == (pool_size, head_dim)
assert tail_slot_mapping.shape == (num_requests, next_n)
assert slot_mapping.shape == (num_requests, next_n)
assert positions.shape == (num_requests, next_n)
assert key.dtype == torch.bfloat16
assert slot_score.dtype == torch.bfloat16
assert ape.dtype == torch.float32
assert kv_cache.dtype == torch.uint8
page_size = kv_cache.shape[1]
buf = kv_cache
buf_fp8 = buf.view(FP8_DTYPE)
buf_fp32 = buf.view(torch.float32)
# The kernel indexes the int tensors as ``req * next_n + t`` (row-major),
# so they must be contiguous. Callers pass either a view of a contiguous
# slice or a freshly scattered tensor, making these no-ops; the calls guard
# against a future caller handing over a strided view.
tail_slot_mapping = tail_slot_mapping.contiguous()
slot_mapping = slot_mapping.contiguous()
positions = positions.contiguous()
if page_size > 1:
assert page_size % 16 == 0, "ROCm preshuffle requires 16-token tiles"
_kpool_decode_update_batched_kernel[(num_requests,)](
buf_fp8,
buf_fp32,
tail_kv_cache,
tail_slot_mapping,
key,
key.stride(0),
key.stride(1),
slot_score,
slot_score.stride(0),
slot_score.stride(1),
ape,
ape.stride(0),
slot_mapping,
positions,
next_n,
PAGE_SIZE=page_size,
BUF_NUMEL_PER_PAGE=buf.stride(0),
POOL_SIZE=pool_size,
TAIL_BLOCK_ELEMS=tail_kv_cache.stride(0),
KPOOL_HEAD=tail_kv_cache.stride(1),
HEAD_DIM=head_dim,
S_OFFSET_NBYTES_IN_PAGE=page_size * head_dim,
FP8_MAX=FP8_MAX,
PRESHUFFLE=page_size > 1,
ROUND_SCALE=round_scale,
BLOCK_D=triton.next_power_of_2(head_dim),
)
# Pool-level top-k helpers.
def history_group_budget_for_topk(topk: int, pool_size: int) -> int:
"""Number of pools to select so that expanding yields ``topk`` tokens."""
assert topk % pool_size == 0
return topk // pool_size
def expand_pools_to_tokens(
group_ids: torch.Tensor,
group_valid: torch.Tensor,
topk: int,
pool_size: int,
page_table: torch.Tensor | None = None,
topk_offsets: torch.Tensor | None = None,
) -> torch.Tensor:
"""Expand selected full-pool ids to a strict-width token topk tensor."""
assert group_ids.ndim == 2
assert group_valid.shape == group_ids.shape
assert topk % pool_size == 0
assert group_ids.shape[1] == history_group_budget_for_topk(topk, pool_size)
assert page_table is None or topk_offsets is None
device = group_ids.device
offsets = torch.arange(pool_size, device=device, dtype=torch.int64)
token_ids = group_ids.to(torch.int64).unsqueeze(-1) * pool_size + offsets
token_ids = token_ids.reshape(group_ids.shape[0], topk)
valid = (
group_valid.unsqueeze(-1)
.expand(-1, -1, pool_size)
.reshape(group_ids.shape[0], topk)
)
if page_table is not None:
assert page_table.ndim == 2
safe_ids = token_ids.clamp(min=0, max=page_table.shape[1] - 1)
output = torch.gather(page_table, dim=1, index=safe_ids).to(torch.int32)
elif topk_offsets is not None:
if topk_offsets.ndim == 2:
assert topk_offsets.shape[1] == 1
topk_offsets = topk_offsets.squeeze(1)
output = (token_ids + topk_offsets.to(torch.int64).unsqueeze(1)).to(torch.int32)
else:
output = token_ids.to(torch.int32)
return torch.where(valid, output, torch.full_like(output, -1))
def append_tail_to_topk(
topk_result: torch.Tensor,
seq_lens: torch.Tensor,
pool_lens: torch.Tensor,
pool_size: int,
page_table: torch.Tensor | None = None,
topk_offsets: torch.Tensor | None = None,
) -> torch.Tensor:
"""Append non-pooled tail tokens after expanded history tokens.
``index_kpool_always_select_tail`` keeps the (incomplete) trailing pool so
the most recent tokens are always attended to.
"""
assert topk_result.dtype == torch.int32
assert seq_lens.ndim == 1
assert pool_lens.ndim == 1
tail_pool = pool_size - 1
if tail_pool == 0:
return topk_result
rows, n_cols = topk_result.shape
out_cols = n_cols + tail_pool
out = torch.empty(
(rows, out_cols), dtype=topk_result.dtype, device=topk_result.device
)
# tail tokens: [pool_len*pool_size, seq_len) for each row.
pool_len = pool_lens.to(torch.int32)
tail_start = pool_len * pool_size
seq_len = seq_lens.to(torch.int32)
tail_count = seq_len - tail_start # in [0, pool_size)
cols = torch.arange(out_cols, device=topk_result.device)[None, :]
history_len = n_cols
is_history = cols < history_len
tail_off = cols - history_len
is_tail = (tail_off >= 0) & (tail_off < tail_count[:, None])
# safe_hist must be per-row [rows, out_cols] so the gather reads each row's
# OWN history. cols is [1, out_cols]; if used directly, gather (which does
# NOT broadcast the index) would read only row 0 of topk_result, making every
# query inherit row 0's history (empty for the first token) and lose all its
# selected tokens — only the per-row tail would survive. This only manifests
# for multi-row sparse PREFILL (decode has 1 row, so it reads its own row 0).
safe_hist = torch.minimum(cols, torch.full_like(cols, n_cols - 1)).expand(
rows, out_cols
)
history_val = torch.gather(topk_result, 1, safe_hist)
tail_raw = tail_start[:, None] + tail_off
tail_val = tail_raw.to(torch.int32)
if page_table is not None:
safe_tail = tail_raw.clamp(min=0, max=page_table.shape[1] - 1)
tail_val = torch.gather(page_table, 1, safe_tail).to(torch.int32)
elif topk_offsets is not None:
tail_val = (tail_raw + topk_offsets.to(torch.int64).unsqueeze(1)).to(
torch.int32
)
out = torch.where(is_history, history_val, -1)
out = torch.where(is_tail, tail_val, out)
return out
@triton.jit
def _expand_pools_and_append_tail_kernel(
pool_ids_ptr, # [rows, n_groups], int (any int dtype)
seq_lens_ptr, # [rows], int32 (token-granular seq_len)
out_ptr, # [rows, out_cols], int32
topk, # n_groups * pool_size
out_cols, # topk + pool_size - 1
POOL_SIZE: tl.constexpr,
BLOCK_COLS: tl.constexpr,
pid_s0,
out_s0,
):
# Fuses expand_pools_to_tokens + append_tail_to_topk (identity path) into a
# single kernel. Each program writes one (row, column-tile) of the output.
row = tl.program_id(0)
tile = tl.program_id(1)
cols = tile * BLOCK_COLS + tl.arange(0, BLOCK_COLS)
mask = cols < out_cols
seq_len = tl.load(seq_lens_ptr + row)
pool_len = seq_len // POOL_SIZE
tail_start = pool_len * POOL_SIZE
tail_count = seq_len - tail_start # in [0, POOL_SIZE)
# History region [0, topk): expand selected pool g = cols // POOL_SIZE.
is_history = cols < topk
g = cols // POOL_SIZE
o = cols % POOL_SIZE
pid = tl.load(pool_ids_ptr + row * pid_s0 + g, mask=mask & is_history, other=-1)
hist_val = (pid * POOL_SIZE + o).to(tl.int32)
hist_out = tl.where(pid >= 0, hist_val, -1)
# Tail region [topk, out_cols): the request's trailing incomplete pool.
tail_off = cols - topk
is_tail = (tail_off >= 0) & (tail_off < tail_count)
tail_val = (tail_start + tail_off).to(tl.int32)
tail_out = tl.where(is_tail, tail_val, -1)
result = tl.where(is_history, hist_out, tail_out)
tl.store(out_ptr + row * out_s0 + cols, result, mask=mask)
def expand_pools_and_append_tail(
pool_ids: torch.Tensor,
seq_lens: torch.Tensor,
pool_size: int,
) -> torch.Tensor:
"""Fuse ``expand_pools_to_tokens`` + ``append_tail_to_topk`` (identity path).
Produces the same ``[rows, topk + pool_size - 1]`` int32 output as calling
the two functions in sequence when neither ``page_table`` nor
``topk_offsets`` is passed — the only path used by the GLM-5.3-Flash indexer.
The kernel derives ``pool_len = seq_len // pool_size`` internally, so the
caller no longer needs to precompute it. Replaces ~25 elementwise kernels
with one Triton launch.
"""
rows, n_groups = pool_ids.shape
topk = n_groups * pool_size
out_cols = topk + pool_size - 1
out = torch.empty((rows, out_cols), dtype=torch.int32, device=pool_ids.device)
BLOCK_COLS = 128
n_tiles = triton.cdiv(out_cols, BLOCK_COLS)
_expand_pools_and_append_tail_kernel[(rows, n_tiles)](
pool_ids,
seq_lens,
out,
topk,
out_cols,
POOL_SIZE=pool_size,
BLOCK_COLS=BLOCK_COLS,
pid_s0=pool_ids.stride(0),
out_s0=out.stride(0),
)
return out
+2
View File
@@ -0,0 +1,2 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
@@ -0,0 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from .kernels import chunk_kda_with_fused_gate, fused_recurrent_kda
__all__ = ["chunk_kda_with_fused_gate", "fused_recurrent_kda"]
@@ -0,0 +1,656 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
# mypy: ignore-errors
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501
import torch
from vllm.third_party.flash_linear_attention.ops.op import exp
from vllm.triton_utils import tl, triton
@triton.heuristics(
{
"USE_INITIAL_STATE": lambda args: args["h0"] is not None,
"IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
"IS_CONTINUOUS_BATCHING": lambda args: args["ssm_state_indices"] is not None,
"IS_SPEC_DECODING": lambda args: args["num_accepted_tokens"] is not None,
}
)
@triton.jit(do_not_specialize=["N", "T"])
def fused_recurrent_gated_delta_rule_fwd_kernel(
q,
k,
v,
g,
beta,
o,
h0,
ht,
cu_seqlens,
ssm_state_indices,
num_accepted_tokens,
a_log,
g_bias,
scale,
N: tl.int64, # num of sequences
T: tl.int64, # num of tokens
B: tl.constexpr,
H: tl.constexpr,
HV: tl.constexpr,
K: tl.constexpr,
V: tl.constexpr,
BK: tl.constexpr,
BV: tl.constexpr,
stride_init_state_token: tl.constexpr,
stride_final_state_token: tl.constexpr,
stride_indices_seq: tl.constexpr,
stride_indices_tok: tl.constexpr,
USE_INITIAL_STATE: tl.constexpr, # whether to use initial state
INPLACE_FINAL_STATE: tl.constexpr, # whether to store final state inplace
IS_BETA_HEADWISE: tl.constexpr, # whether beta is headwise vector or scalar,
USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
IS_VARLEN: tl.constexpr,
IS_CONTINUOUS_BATCHING: tl.constexpr,
IS_SPEC_DECODING: tl.constexpr,
IS_KDA: tl.constexpr,
SIGMOID_BETA: tl.constexpr, # beta holds raw logits; sigmoid at fp32 load
COMPUTE_GATE: tl.constexpr, # g holds raw logits; KDA gate computed in-kernel
SAFE_GATE: tl.constexpr, # bounded gate variant (only branch implemented)
LOWER_BOUND: tl.constexpr,
):
i_k, i_v, i_nh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
i_n, i_hv = i_nh // HV, i_nh % HV
i_h = i_hv // (HV // H)
if IS_VARLEN:
bos, eos = (
tl.load(cu_seqlens + i_n).to(tl.int64),
tl.load(cu_seqlens + i_n + 1).to(tl.int64),
)
all = T
T = eos - bos
else:
bos, eos = i_n * T, i_n * T + T
all = B * T
if T == 0:
# no tokens to process for this sequence
return
o_k = i_k * BK + tl.arange(0, BK)
o_v = i_v * BV + tl.arange(0, BV)
p_q = q + (bos * H + i_h) * K + o_k
p_k = k + (bos * H + i_h) * K + o_k
p_v = v + (bos * HV + i_hv) * V + o_v
if IS_BETA_HEADWISE:
p_beta = beta + (bos * HV + i_hv) * V + o_v
else:
p_beta = beta + bos * HV + i_hv
if not IS_KDA:
p_g = g + bos * HV + i_hv
else:
p_gk = g + (bos * HV + i_hv) * K + o_k
# Per-head gate amplitude, hoisted out of the token loop (COMPUTE_GATE).
if COMPUTE_GATE:
b_a_log = tl.exp(tl.load(a_log + i_h).to(tl.float32))
p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v
mask_k = o_k < K
mask_v = o_v < V
mask_h = mask_v[:, None] & mask_k[None, :]
b_h = tl.zeros([BV, BK], dtype=tl.float32)
if USE_INITIAL_STATE:
if IS_CONTINUOUS_BATCHING:
if IS_SPEC_DECODING:
i_t = tl.load(num_accepted_tokens + i_n).to(tl.int64) - 1
else:
i_t = 0
# Load state index and check for invalid entries
state_idx = tl.load(ssm_state_indices + i_n * stride_indices_seq + i_t).to(
tl.int64
)
# Skip if state index is invalid (NULL_BLOCK_ID=0)
if state_idx <= 0:
return
p_h0 = h0 + state_idx * stride_init_state_token
else:
p_h0 = h0 + bos * HV * V * K
p_h0 = p_h0 + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
b_h += tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)
for i_t in range(0, T):
b_q = tl.load(p_q, mask=mask_k, other=0).to(tl.float32)
b_k = tl.load(p_k, mask=mask_k, other=0).to(tl.float32)
b_v = tl.load(p_v, mask=mask_v, other=0).to(tl.float32)
if USE_QK_L2NORM_IN_KERNEL:
b_q = b_q / tl.sqrt(tl.sum(b_q * b_q) + 1e-6)
b_k = b_k / tl.sqrt(tl.sum(b_k * b_k) + 1e-6)
b_q = b_q * scale
# [BV, BK]
if not IS_KDA:
b_g = tl.load(p_g).to(tl.float32)
b_h *= exp(b_g)
else:
b_gk = tl.load(p_gk).to(tl.float32)
if COMPUTE_GATE:
# Replicates kda_gate_fwd_kernel's SAFE_GATE branch
# bit-for-bit (same tl.exp, same fp32 math; the intermediate
# gate value this replaces was stored/reloaded as fp32,
# which is lossless): y = lb / (1 + exp(-exp(A)*(g+bias))).
b_gk += tl.load(g_bias + i_h * K + o_k, mask=mask_k, other=0.0).to(
tl.float32
)
b_gk = LOWER_BOUND / (1.0 + tl.exp(-(b_a_log * b_gk)))
b_h *= exp(b_gk[None, :])
# [BV]
b_v -= tl.sum(b_h * b_k[None, :], 1)
if IS_BETA_HEADWISE:
b_beta = tl.load(p_beta, mask=mask_v, other=0).to(tl.float32)
else:
b_beta = tl.load(p_beta).to(tl.float32)
# Matches torch's `x.float().sigmoid()` pre-computation bit-for-bit
# on the input side (bf16->fp32 is exact); only the sigmoid impl itself
# can differ by <=1 ULP.
if SIGMOID_BETA:
b_beta = tl.sigmoid(b_beta)
b_v *= b_beta
# [BV, BK]
b_h += b_v[:, None] * b_k[None, :]
# [BV]
b_o = tl.sum(b_h * b_q[None, :], 1)
tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)
# keep the states for multi-query tokens
if INPLACE_FINAL_STATE:
# Load state index and check for invalid entries
final_state_idx = tl.load(
ssm_state_indices + i_n * stride_indices_seq + i_t
).to(tl.int64)
# Only store if state index is valid (not NULL_BLOCK_ID=0)
if final_state_idx > 0:
p_ht = ht + final_state_idx * stride_final_state_token
p_ht = p_ht + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
else:
p_ht = ht + (bos + i_t) * stride_final_state_token
p_ht = p_ht + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
p_q += H * K
p_k += H * K
p_o += HV * V
p_v += HV * V
if not IS_KDA:
p_g += HV
else:
p_gk += HV * K
p_beta += HV * (V if IS_BETA_HEADWISE else 1)
def fused_recurrent_gated_delta_rule_fwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
inplace_final_state: bool = True,
cu_seqlens: torch.Tensor | None = None,
ssm_state_indices: torch.Tensor | None = None,
num_accepted_tokens: torch.Tensor | None = None,
use_qk_l2norm_in_kernel: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
B, T, H, K, V = *k.shape, v.shape[-1]
HV = v.shape[2]
N = B if cu_seqlens is None else len(cu_seqlens) - 1
BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 32)
NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
assert NK == 1, "NK > 1 is not supported yet"
num_stages = 3
num_warps = 1
o = q.new_empty(NK, *v.shape)
if inplace_final_state:
final_state = initial_state
else:
final_state = q.new_empty(T, HV, V, K, dtype=initial_state.dtype)
stride_init_state_token = initial_state.stride(0)
stride_final_state_token = final_state.stride(0)
if ssm_state_indices is None:
stride_indices_seq, stride_indices_tok = 1, 1
elif ssm_state_indices.ndim == 1:
stride_indices_seq, stride_indices_tok = ssm_state_indices.stride(0), 1
else:
stride_indices_seq, stride_indices_tok = ssm_state_indices.stride()
grid = (NK, NV, N * HV)
fused_recurrent_gated_delta_rule_fwd_kernel[grid](
q=q,
k=k,
v=v,
g=g,
beta=beta,
o=o,
h0=initial_state,
ht=final_state,
cu_seqlens=cu_seqlens,
ssm_state_indices=ssm_state_indices,
num_accepted_tokens=num_accepted_tokens,
scale=scale,
N=N,
T=T,
B=B,
H=H,
HV=HV,
K=K,
V=V,
BK=BK,
BV=BV,
stride_init_state_token=stride_init_state_token,
stride_final_state_token=stride_final_state_token,
stride_indices_seq=stride_indices_seq,
stride_indices_tok=stride_indices_tok,
IS_BETA_HEADWISE=beta.ndim == v.ndim,
USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
INPLACE_FINAL_STATE=inplace_final_state,
IS_KDA=False,
SIGMOID_BETA=False,
a_log=None,
g_bias=None,
COMPUTE_GATE=False,
SAFE_GATE=True,
LOWER_BOUND=-5.0,
num_warps=num_warps,
num_stages=num_stages,
)
o = o.squeeze(0)
return o, final_state
@triton.jit
def fused_recurrent_gated_delta_rule_packed_decode_kernel(
mixed_qkv,
a,
b,
A_log,
dt_bias,
o,
h0,
ht,
ssm_state_indices,
scale,
stride_mixed_qkv_tok: tl.constexpr,
stride_a_tok: tl.constexpr,
stride_b_tok: tl.constexpr,
stride_init_state_token: tl.constexpr,
stride_final_state_token: tl.constexpr,
stride_indices_seq: tl.constexpr,
H: tl.constexpr,
HV: tl.constexpr,
K: tl.constexpr,
V: tl.constexpr,
BK: tl.constexpr,
BV: tl.constexpr,
SOFTPLUS_THRESHOLD: tl.constexpr,
USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
SPLIT_BATCH_HEAD_GRID: tl.constexpr,
):
if SPLIT_BATCH_HEAD_GRID:
i_v, i_hv, i_n = tl.program_id(0), tl.program_id(1), tl.program_id(2)
else:
i_v, i_nh = tl.program_id(0), tl.program_id(1)
i_n, i_hv = i_nh // HV, i_nh % HV
i_h = i_hv // (HV // H)
o_k = tl.arange(0, BK)
o_v = i_v * BV + tl.arange(0, BV)
mask_k = o_k < K
mask_v = o_v < V
mask_h = mask_v[:, None] & mask_k[None, :]
state_idx = tl.load(ssm_state_indices + i_n * stride_indices_seq).to(tl.int64)
p_o = o + (i_n * HV + i_hv) * V + o_v
# Skip if state index is invalid (NULL_BLOCK_ID=0)
if state_idx <= 0:
zero = tl.zeros([BV], dtype=tl.float32).to(p_o.dtype.element_ty)
tl.store(p_o, zero, mask=mask_v)
return
p_h0 = h0 + state_idx * stride_init_state_token
p_h0 = p_h0 + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
b_h = tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)
p_mixed = mixed_qkv + i_n * stride_mixed_qkv_tok
q_off = i_h * K + o_k
k_off = (H * K) + i_h * K + o_k
v_off = (2 * H * K) + i_hv * V + o_v
b_q = tl.load(p_mixed + q_off, mask=mask_k, other=0).to(tl.float32)
b_k = tl.load(p_mixed + k_off, mask=mask_k, other=0).to(tl.float32)
b_v = tl.load(p_mixed + v_off, mask=mask_v, other=0).to(tl.float32)
if USE_QK_L2NORM_IN_KERNEL:
b_q = b_q / tl.sqrt(tl.sum(b_q * b_q) + 1e-6)
b_k = b_k / tl.sqrt(tl.sum(b_k * b_k) + 1e-6)
b_q = b_q * scale
a_val = tl.load(a + i_n * stride_a_tok + i_hv).to(tl.float32)
b_val = tl.load(b + i_n * stride_b_tok + i_hv).to(tl.float32)
A_log_val = tl.load(A_log + i_hv).to(tl.float32)
dt_bias_val = tl.load(dt_bias + i_hv).to(tl.float32)
x = a_val + dt_bias_val
softplus_x = tl.where(x <= SOFTPLUS_THRESHOLD, tl.log(1.0 + tl.exp(x)), x)
g_val = -tl.exp(A_log_val) * softplus_x
beta_val = tl.sigmoid(b_val)
b_h *= exp(g_val)
b_v -= tl.sum(b_h * b_k[None, :], 1)
b_v *= beta_val
b_h += b_v[:, None] * b_k[None, :]
b_o = tl.sum(b_h * b_q[None, :], 1)
tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)
p_ht = ht + state_idx * stride_final_state_token
p_ht = p_ht + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
def fused_recurrent_gated_delta_rule_packed_decode(
mixed_qkv: torch.Tensor,
a: torch.Tensor,
b: torch.Tensor,
A_log: torch.Tensor,
dt_bias: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
out: torch.Tensor,
ssm_state_indices: torch.Tensor,
use_qk_l2norm_in_kernel: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
if mixed_qkv.ndim != 2:
raise ValueError(
f"`mixed_qkv` must be a 2D tensor (got ndim={mixed_qkv.ndim})."
)
if mixed_qkv.stride(-1) != 1:
raise ValueError("`mixed_qkv` must be contiguous in the last dim.")
if a.ndim != 2 or b.ndim != 2:
raise ValueError(
f"`a` and `b` must be 2D tensors (got a.ndim={a.ndim}, b.ndim={b.ndim})."
)
if a.stride(-1) != 1 or b.stride(-1) != 1:
raise ValueError("`a`/`b` must be contiguous in the last dim.")
if A_log.ndim != 1 or dt_bias.ndim != 1:
raise ValueError("`A_log`/`dt_bias` must be 1D tensors.")
if A_log.stride(0) != 1 or dt_bias.stride(0) != 1:
raise ValueError("`A_log`/`dt_bias` must be contiguous.")
if ssm_state_indices.ndim != 1:
raise ValueError(
f"`ssm_state_indices` must be 1D for packed decode (got ndim={ssm_state_indices.ndim})."
)
if not out.is_contiguous():
raise ValueError("`out` must be contiguous.")
dev = mixed_qkv.device
if (
a.device != dev
or b.device != dev
or A_log.device != dev
or dt_bias.device != dev
or initial_state.device != dev
or out.device != dev
or ssm_state_indices.device != dev
):
raise ValueError("All inputs must be on the same device.")
B = mixed_qkv.shape[0]
if a.shape[0] != B or b.shape[0] != B:
raise ValueError(
"Mismatched batch sizes: "
f"mixed_qkv.shape[0]={B}, a.shape[0]={a.shape[0]}, b.shape[0]={b.shape[0]}."
)
if ssm_state_indices.shape[0] != B:
raise ValueError(
f"`ssm_state_indices` must have shape [B] (got {tuple(ssm_state_indices.shape)}; expected ({B},))."
)
if initial_state.ndim != 4:
raise ValueError(
f"`initial_state` must be a 4D tensor (got ndim={initial_state.ndim})."
)
if initial_state.stride(-1) != 1:
raise ValueError("`initial_state` must be contiguous in the last dim.")
HV, V, K = initial_state.shape[-3:]
if a.shape[1] != HV or b.shape[1] != HV:
raise ValueError(
f"`a`/`b` must have shape [B, HV] with HV={HV} (got a.shape={tuple(a.shape)}, b.shape={tuple(b.shape)})."
)
if A_log.numel() != HV or dt_bias.numel() != HV:
raise ValueError(
f"`A_log` and `dt_bias` must have {HV} elements (got A_log.numel()={A_log.numel()}, dt_bias.numel()={dt_bias.numel()})."
)
if out.shape != (B, 1, HV, V):
raise ValueError(
f"`out` must have shape {(B, 1, HV, V)} (got out.shape={tuple(out.shape)})."
)
qkv_dim = mixed_qkv.shape[1]
qk_dim = qkv_dim - HV * V
if qk_dim <= 0 or qk_dim % 2 != 0:
raise ValueError(
f"Invalid packed `mixed_qkv` last dim={qkv_dim} for HV={HV}, V={V}."
)
q_dim = qk_dim // 2
if q_dim % K != 0:
raise ValueError(f"Invalid packed Q size {q_dim}: must be divisible by K={K}.")
H = q_dim // K
if H <= 0 or HV % H != 0:
raise ValueError(
f"Invalid head config inferred from mixed_qkv: H={H}, HV={HV}."
)
BK = triton.next_power_of_2(K)
if triton.cdiv(K, BK) != 1:
raise ValueError(
f"Packed decode kernel only supports NK=1 (got K={K}, BK={BK})."
)
BV = min(triton.next_power_of_2(V), 32)
num_stages = 3
num_warps = 1
stride_mixed_qkv_tok = mixed_qkv.stride(0)
stride_a_tok = a.stride(0)
stride_b_tok = b.stride(0)
stride_init_state_token = initial_state.stride(0)
stride_final_state_token = initial_state.stride(0)
stride_indices_seq = ssm_state_indices.stride(0)
NV = triton.cdiv(V, BV)
# CUDA limits grid Y/Z dimensions to 65535.
split_batch_head_grid = B * HV > 65535
grid = (NV, HV, B) if split_batch_head_grid else (NV, B * HV)
fused_recurrent_gated_delta_rule_packed_decode_kernel[grid](
mixed_qkv=mixed_qkv,
a=a,
b=b,
A_log=A_log,
dt_bias=dt_bias,
o=out,
h0=initial_state,
ht=initial_state,
ssm_state_indices=ssm_state_indices,
scale=scale,
stride_mixed_qkv_tok=stride_mixed_qkv_tok,
stride_a_tok=stride_a_tok,
stride_b_tok=stride_b_tok,
stride_init_state_token=stride_init_state_token,
stride_final_state_token=stride_final_state_token,
stride_indices_seq=stride_indices_seq,
H=H,
HV=HV,
K=K,
V=V,
BK=BK,
BV=BV,
SOFTPLUS_THRESHOLD=20.0,
USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
SPLIT_BATCH_HEAD_GRID=split_batch_head_grid,
num_warps=num_warps,
num_stages=num_stages,
)
return out, initial_state
class FusedRecurrentFunction(torch.autograd.Function):
@staticmethod
def forward(
ctx,
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
inplace_final_state: bool = True,
cu_seqlens: torch.Tensor | None = None,
ssm_state_indices: torch.Tensor | None = None,
num_accepted_tokens: torch.Tensor | None = None,
use_qk_l2norm_in_kernel: bool = False,
):
o, final_state = fused_recurrent_gated_delta_rule_fwd(
q=q.contiguous(),
k=k.contiguous(),
v=v.contiguous(),
g=g.contiguous(),
beta=beta.contiguous(),
scale=scale,
initial_state=initial_state,
inplace_final_state=inplace_final_state,
cu_seqlens=cu_seqlens,
ssm_state_indices=ssm_state_indices,
num_accepted_tokens=num_accepted_tokens,
use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
)
return o, final_state
def fused_recurrent_gated_delta_rule(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor = None,
scale: float = None,
initial_state: torch.Tensor = None,
inplace_final_state: bool = True,
cu_seqlens: torch.Tensor | None = None,
ssm_state_indices: torch.Tensor | None = None,
num_accepted_tokens: torch.Tensor | None = None,
use_qk_l2norm_in_kernel: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
r"""
Args:
q (torch.Tensor):
queries of shape `[B, T, H, K]`.
k (torch.Tensor):
keys of shape `[B, T, H, K]`.
v (torch.Tensor):
values of shape `[B, T, HV, V]`.
GVA is applied if `HV > H`.
g (torch.Tensor):
g (decays) of shape `[B, T, HV]`.
beta (torch.Tensor):
betas of shape `[B, T, HV]`.
scale (Optional[int]):
Scale factor for the RetNet attention scores.
If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
initial_state (Optional[torch.Tensor]):
Initial state of shape `[N, HV, V, K]` for `N` input sequences.
For equal-length input sequences, `N` equals the batch size `B`.
Default: `None`.
inplace_final_state: bool:
Whether to store the final state in-place to save memory.
Default: `True`.
cu_seqlens (torch.Tensor):
Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
consistent with the FlashAttention API.
ssm_state_indices (Optional[torch.Tensor]):
Indices to map the input sequences to the initial/final states.
num_accepted_tokens (Optional[torch.Tensor]):
Number of accepted tokens for each sequence during decoding.
Returns:
o (torch.Tensor):
Outputs of shape `[B, T, HV, V]`.
final_state (torch.Tensor):
Final state of shape `[N, HV, V, K]`.
Examples::
>>> import torch
>>> import torch.nn.functional as F
>>> from einops import rearrange
>>> from fla.ops.gated_delta_rule import fused_recurrent_gated_delta_rule
# inputs with equal lengths
>>> B, T, H, HV, K, V = 4, 2048, 4, 8, 512, 512
>>> q = torch.randn(B, T, H, K, device='cuda')
>>> k = F.normalize(torch.randn(B, T, H, K, device='cuda'), p=2, dim=-1)
>>> v = torch.randn(B, T, HV, V, device='cuda')
>>> g = F.logsigmoid(torch.rand(B, T, HV, device='cuda'))
>>> beta = torch.rand(B, T, HV, device='cuda').sigmoid()
>>> h0 = torch.randn(B, HV, V, K, device='cuda')
>>> o, ht = fused_gated_recurrent_delta_rule(
q, k, v, g, beta,
initial_state=h0,
)
# for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
>>> q, k, v, g, beta = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, g, beta))
# for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
>>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.int32)
>>> o_var, ht_var = fused_gated_recurrent_delta_rule(
q, k, v, g, beta,
initial_state=h0,
cu_seqlens=cu_seqlens
)
"""
if cu_seqlens is not None and q.shape[0] != 1:
raise ValueError(
f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`."
f"Please flatten variable-length inputs before processing."
)
if scale is None:
scale = k.shape[-1] ** -0.5
else:
assert scale > 0, "scale must be positive"
if beta is None:
beta = torch.ones_like(q[..., 0])
o, final_state = FusedRecurrentFunction.apply(
q,
k,
v,
g,
beta,
scale,
initial_state,
inplace_final_state,
cu_seqlens,
ssm_state_indices,
num_accepted_tokens,
use_qk_l2norm_in_kernel,
)
return o, final_state
File diff suppressed because it is too large Load Diff