[Kernel][Build] Bump DeepGEMM to the 26/09/30 release and adapt the sparse indexer (#61149)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-10-12 23:24:43 -07:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 86b1c2ef15
commit a6009ea85b
23 changed files with 516 additions and 208 deletions
@@ -15,6 +15,8 @@ from __future__ import annotations
import torch
from vllm.triton_utils import tl, triton
from vllm.utils.deep_gemm import mqa_logits_use_mx_scales
from vllm.v1.attention.ops.common import pack_ue8m0_x4
# The GLM-5.3-Flash indexer head dimension is fixed at 128.
INDEX_HEAD_DIM = 128
@@ -157,6 +159,7 @@ def _kpool_softmax_rotate_write_cache_kernel(
HEAD_DIM: tl.constexpr,
S_OFFSET_NBYTES_IN_PAGE: tl.constexpr,
ROUND_SCALE: tl.constexpr,
MX_SCALE: tl.constexpr,
HAS_WRITE_MASK: tl.constexpr,
RETURN_COMPRESSED: tl.constexpr,
WRITE_CACHE: tl.constexpr,
@@ -245,7 +248,14 @@ def _kpool_softmax_rotate_write_cache_kernel(
+ 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 MX_SCALE:
tl.store(
buf_fp32_ptr.to(tl.pointer_type(tl.int32)) + out_s_offset,
pack_ue8m0_x4(scale),
mask=do_write,
)
else:
tl.store(buf_fp32_ptr + out_s_offset, scale, mask=do_write)
if RETURN_COMPRESSED:
tl.store(
@@ -324,6 +334,7 @@ def kpool_compress_and_write_cache(
buf_fp8 = buf.view(torch.float8_e4m3fn)
buf_fp32 = buf.view(torch.float32)
mx_scale = write_cache and mqa_logits_use_mx_scales()
# 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
@@ -361,7 +372,9 @@ def kpool_compress_and_write_cache(
POOL_SIZE=slot_k.shape[1],
HEAD_DIM=head_dim,
S_OFFSET_NBYTES_IN_PAGE=s_offset_nbytes_in_page,
ROUND_SCALE=round_scale,
# MX scales need power-of-two scales.
ROUND_SCALE=round_scale or mx_scale,
MX_SCALE=mx_scale,
HAS_WRITE_MASK=has_write_mask,
RETURN_COMPRESSED=return_compressed,
WRITE_CACHE=write_cache,
@@ -479,6 +492,7 @@ def _kpool_decode_update_batched_kernel(
HEAD_DIM: tl.constexpr,
S_OFFSET_NBYTES_IN_PAGE: tl.constexpr,
ROUND_SCALE: tl.constexpr,
MX_SCALE: tl.constexpr,
BLOCK_D: tl.constexpr,
):
"""One program per request; iterates its NEXT_N verify tokens in order.
@@ -604,7 +618,13 @@ def _kpool_decode_update_batched_kernel(
+ 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)
if MX_SCALE:
tl.store(
buf_fp32_ptr.to(tl.pointer_type(tl.int32)) + out_s_offset,
pack_ue8m0_x4(scale),
)
else:
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
@@ -718,7 +738,9 @@ def kpool_decode_update_and_maybe_write_cache_batched(
KPOOL_HEAD=tail_kv_cache.stride(1),
HEAD_DIM=head_dim,
S_OFFSET_NBYTES_IN_PAGE=page_size * head_dim,
ROUND_SCALE=round_scale,
# MX scales need power-of-two scales.
ROUND_SCALE=round_scale or mqa_logits_use_mx_scales(),
MX_SCALE=mqa_logits_use_mx_scales(),
BLOCK_D=triton.next_power_of_2(head_dim),
)
+46 -26
View File
@@ -16,6 +16,7 @@ from vllm.model_executor.layers.sparse_attn_indexer import (
PrefillRowShard,
_merge_dcp_topk_global,
_reserve_row_shard_exchange_transient,
as_fp32_logits,
)
from vllm.models.glm5next.common.sparse_indexer import (
RADIX_TOPK_WORKSPACE_SIZE,
@@ -29,7 +30,13 @@ from vllm.models.glm5next.common.sparse_indexer import (
)
from vllm.models.glm5next.nvidia.ops import kpool_compress as kpool_ops
from vllm.platforms import current_platform
from vllm.utils.deep_gemm import has_deep_gemm
from vllm.utils.deep_gemm import (
fp8_fp4_mqa_logits,
fp8_fp4_paged_mqa_logits,
has_deep_gemm,
mqa_logits_mx_inputs,
mqa_logits_use_mx_scales,
)
from vllm.utils.import_utils import has_cutedsl
from vllm.utils.torch_utils import (
LayerNameType,
@@ -253,7 +260,7 @@ def sparse_attn_indexer_kpool(
kv_cache,
slot_mapping,
quant_block_size,
scale_fmt,
"mxfp8" if mqa_logits_use_mx_scales() else scale_fmt,
)
topk_indices_buffer[: hidden_states.shape[0]] = -1
@@ -343,20 +350,29 @@ def sparse_attn_indexer_kpool(
else:
q_slice_cast = q_slice
k_quant_cast = k_quant
k_scale_cast = k_scale.view(torch.float32).squeeze(-1)
from vllm.utils.deep_gemm import fp8_fp4_mqa_logits
k_scale_cast = k_scale.view(
torch.int32 if mqa_logits_use_mx_scales() else torch.float32
).squeeze(-1)
num_rows = q_slice.shape[0]
if chunk.local_total_seq_lens == 0:
logits = q_slice.new_empty((num_rows, 0), dtype=torch.float32)
else:
logits = fp8_fp4_mqa_logits(
(q_slice_cast, q_scale_slice),
(k_quant_cast, k_scale_cast),
weights[row_start:row_end],
cu_seqlen_ks,
cu_seqlen_ke,
clean_logits=False,
q_scale_slice, chunk_weights = mqa_logits_mx_inputs(
q_slice_cast, q_scale_slice, weights[row_start:row_end]
)
# Compressed: row i starts at K position cu_seqlen_ks[i].
row_ends = cu_seqlen_ke - cu_seqlen_ks
logits = as_fp32_logits(
fp8_fp4_mqa_logits(
(q_slice_cast, q_scale_slice),
(k_quant_cast, k_scale_cast),
chunk_weights,
cu_seqlen_ks,
cu_seqlen_ke,
chunk.max_seqlen_k,
),
row_ends,
)
# kpool: logits are pool-granular (compress_ratio == index_kpool),
@@ -376,8 +392,8 @@ def sparse_attn_indexer_kpool(
else:
torch.ops._C.top_k_per_row_prefill(
logits,
cu_seqlen_ks,
cu_seqlen_ke,
torch.zeros_like(row_ends),
row_ends,
topk_dst,
num_rows,
logits.stride(0),
@@ -392,7 +408,7 @@ def sparse_attn_indexer_kpool(
dcp_rank,
dcp_world_size,
pool_cp_interleave,
row_starts=cu_seqlen_ks,
row_starts=None,
)
if index_kpool > 1:
@@ -594,18 +610,22 @@ def sparse_attn_indexer_kpool(
if use_fp4_cache
else padded_q_quant_decode_tokens
)
from vllm.utils.deep_gemm import fp8_fp4_paged_mqa_logits
logits = fp8_fp4_paged_mqa_logits(
(padded_q_quant_cast, padded_q_scale),
kv_cache,
padded_weights[:num_padded_tokens],
seq_lens,
decode_metadata.block_table,
decode_metadata.schedule_metadata,
max_model_len=max_pool_len,
clean_logits=False,
indices=decode_metadata.indices,
padded_q_scale, decode_weights = mqa_logits_mx_inputs(
padded_q_quant_cast, padded_q_scale, padded_weights[:num_padded_tokens]
)
# SM100 decode is varlen: one row and context length per token.
logits = as_fp32_logits(
fp8_fp4_paged_mqa_logits(
(padded_q_quant_cast, padded_q_scale),
kv_cache,
decode_weights,
seq_lens,
decode_metadata.block_table,
decode_metadata.schedule_metadata,
max_model_len=max_pool_len,
indices=decode_metadata.indices,
),
seq_lens.view(-1),
)
num_rows = logits.shape[0]
# kpool: logits are pool-granular -> select topk_tokens//kpool pools,