[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:
co-authored by
Claude Opus 5.5
parent
86b1c2ef15
commit
a6009ea85b
@@ -6,7 +6,9 @@ import torch
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.triton_utils import tl, triton
|
||||
from vllm.utils.deep_gemm import mqa_logits_use_mx_scales
|
||||
from vllm.utils.torch_utils import is_quantized_kv_cache
|
||||
from vllm.v1.attention.ops.common import pack_ue8m0_x4
|
||||
|
||||
# Cache of tiny 1-element dummy tensors (per device, dtype) reused by the
|
||||
# has_indexer=False path so the indexer args don't allocate every call.
|
||||
@@ -87,6 +89,7 @@ def _fp8_quant_and_cache_write(
|
||||
SHUFFLE: tl.constexpr,
|
||||
BLOCK_TILE: tl.constexpr,
|
||||
HEAD_TILE: tl.constexpr,
|
||||
MX_SCALE: tl.constexpr,
|
||||
FP8_MAX: tl.constexpr,
|
||||
USE_FNUZ: tl.constexpr,
|
||||
):
|
||||
@@ -111,6 +114,8 @@ def _fp8_quant_and_cache_write(
|
||||
|
||||
tl.store(kv_cache_ptr + block_start + value_off, k_fp8, mask=mask)
|
||||
scale_byte_off = block_start + cache_block_size * HEAD_DIM + block_offset * 4
|
||||
if MX_SCALE:
|
||||
scale = pack_ue8m0_x4(scale)
|
||||
tl.store(kv_cache_scale_ptr + scale_byte_off // 4, scale)
|
||||
|
||||
|
||||
@@ -167,6 +172,7 @@ def _fused_norm_rope_kernel(
|
||||
INDEXER_CACHE_SHUFFLE: tl.constexpr,
|
||||
INDEXER_CACHE_BLOCK_TILE: tl.constexpr,
|
||||
INDEXER_CACHE_HEAD_TILE: tl.constexpr,
|
||||
INDEXER_CACHE_MX_SCALE: tl.constexpr,
|
||||
# MLA KV cache (concat kv_c_normed + k_pe_roped, uses slot_mapping_ptr)
|
||||
mla_cache_ptr,
|
||||
mla_cache_block_stride,
|
||||
@@ -501,6 +507,7 @@ def _fused_norm_rope_kernel(
|
||||
INDEXER_CACHE_SHUFFLE,
|
||||
INDEXER_CACHE_BLOCK_TILE,
|
||||
INDEXER_CACHE_HEAD_TILE,
|
||||
INDEXER_CACHE_MX_SCALE,
|
||||
FP8_MAX,
|
||||
USE_FNUZ,
|
||||
)
|
||||
@@ -571,7 +578,10 @@ def fused_norm_rope(
|
||||
assert slot_mapping is not None
|
||||
if indexer_slot_mapping is None:
|
||||
indexer_slot_mapping = slot_mapping
|
||||
idx_cache_scale_view = indexer_k_cache.view(torch.uint8).view(torch.float32)
|
||||
idx_cache_mx_scale = mqa_logits_use_mx_scales()
|
||||
idx_cache_scale_view = indexer_k_cache.view(torch.uint8).view(
|
||||
torch.int32 if idx_cache_mx_scale else torch.float32
|
||||
)
|
||||
idx_cache_block_size = indexer_k_cache.shape[1]
|
||||
idx_cache_stride = indexer_k_cache.shape[2]
|
||||
# The caller's cache reports whether its reader expects the shuffled
|
||||
@@ -591,6 +601,7 @@ def fused_norm_rope(
|
||||
)
|
||||
else:
|
||||
idx_cache_scale_view = None
|
||||
idx_cache_mx_scale = False
|
||||
idx_cache_block_size = 1
|
||||
idx_cache_stride = 0
|
||||
idx_cache_shuffle = False
|
||||
@@ -701,6 +712,7 @@ def fused_norm_rope(
|
||||
idx_cache_shuffle,
|
||||
_INDEXER_CACHE_BLOCK_TILE,
|
||||
idx_cache_head_tile,
|
||||
idx_cache_mx_scale,
|
||||
# MLA KV cache (uses same slot_mapping)
|
||||
mla_kv_cache,
|
||||
mla_block_stride,
|
||||
|
||||
@@ -33,7 +33,9 @@ from vllm.model_executor.warmup.jit_warmup_triton_helper import (
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.triton_utils import tl, triton
|
||||
from vllm.utils.deep_gemm import mqa_logits_use_mx_scales
|
||||
from vllm.utils.math_utils import round_up
|
||||
from vllm.v1.attention.ops.common import pack_ue8m0_x4
|
||||
|
||||
if current_platform.is_rocm():
|
||||
from vllm.platforms.rocm import _ON_GFX950
|
||||
@@ -742,6 +744,7 @@ def _fused_kv_compress_norm_rope_insert_indexer_attn(
|
||||
TOKEN_STRIDE: tl.constexpr, # 128 for indexer
|
||||
SCALE_DIM: tl.constexpr, # 4 for indexer (1 float32)
|
||||
KV_BLOCK_STRIDE: tl.constexpr,
|
||||
MX_SCALE: tl.constexpr, # store the scale as 4 packed UE8M0 bytes
|
||||
):
|
||||
"""Fused compress → RMSNorm → RoPE → FP8 quant → store.
|
||||
|
||||
@@ -878,7 +881,10 @@ def _fused_kv_compress_norm_rope_insert_indexer_attn(
|
||||
|
||||
# Single float32 scale
|
||||
scale_val = tl.exp2(exponent)
|
||||
tl.store(scale_ptr.to(tl.pointer_type(tl.float32)), scale_val)
|
||||
if MX_SCALE:
|
||||
tl.store(scale_ptr.to(tl.pointer_type(tl.int32)), pack_ue8m0_x4(scale_val))
|
||||
else:
|
||||
tl.store(scale_ptr.to(tl.pointer_type(tl.float32)), scale_val)
|
||||
|
||||
|
||||
# =============================================================================
|
||||
@@ -927,6 +933,7 @@ def _fused_kv_compress_norm_rope_insert_indexer_mxfp4_attn(
|
||||
TOKEN_STRIDE: tl.constexpr, # HEAD_SIZE // 2 = 64 packed bytes/token
|
||||
SCALE_DIM: tl.constexpr, # HEAD_SIZE // QUANT_BLOCK = 4 ue8m0 bytes/token
|
||||
KV_BLOCK_STRIDE: tl.constexpr,
|
||||
MX_SCALE: tl.constexpr, # unused for MXFP4 (kept for signature parity)
|
||||
):
|
||||
"""Fused compress → RMSNorm → RoPE → MXFP4 quant → store.
|
||||
|
||||
@@ -1102,6 +1109,7 @@ class FusedKVCompressNormRopeInsertIndexerTritonKernel(
|
||||
block_size: int
|
||||
kv_cache_block_size: int
|
||||
kv_block_stride: int
|
||||
mx_scale: bool
|
||||
|
||||
# Canonical kernel for the launcher's arg_names mapping; the launched
|
||||
# variant (fp8/mxfp4, same arg_names) is picked per call in __call__.
|
||||
@@ -1181,6 +1189,7 @@ class FusedKVCompressNormRopeInsertIndexerTritonKernel(
|
||||
block_size=state_block_size,
|
||||
kv_cache_block_size=kv_cache_block_size,
|
||||
kv_block_stride=kv_block_stride,
|
||||
mx_scale=not use_fp4_cache and mqa_logits_use_mx_scales(),
|
||||
)
|
||||
|
||||
def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]:
|
||||
@@ -1301,6 +1310,7 @@ class FusedKVCompressNormRopeInsertIndexerTritonKernel(
|
||||
TOKEN_STRIDE=token_stride,
|
||||
SCALE_DIM=scale_dim,
|
||||
KV_BLOCK_STRIDE=kv_cache.stride(0),
|
||||
MX_SCALE=not use_fp4_cache and mqa_logits_use_mx_scales(),
|
||||
num_warps=1,
|
||||
**pdl_kwargs,
|
||||
)
|
||||
|
||||
@@ -17,6 +17,8 @@ import torch
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
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
|
||||
|
||||
from . import MXFP4_BLOCK_SIZE, _fp32x2_to_fp4x2
|
||||
|
||||
@@ -51,7 +53,8 @@ def indexer_k_norm_rope_store(
|
||||
kv_slot_mapping: [num_tokens] slots in the indexer cache (-1 = skip).
|
||||
compress_ratio: group size; keys are emitted at group boundaries.
|
||||
use_fp4_cache: MXFP4 (2 nibbles/byte + ue8m0 per 32) when True, else
|
||||
per-token FP8 with a single fp32 scale.
|
||||
per-token FP8 with a single fp32 scale (four packed ue8m0 bytes
|
||||
under `mqa_logits_use_mx_scales`).
|
||||
|
||||
"""
|
||||
num_tokens = kv_slot_mapping.numel()
|
||||
@@ -111,6 +114,7 @@ def indexer_k_norm_rope_store(
|
||||
KV_BLOCK_STRIDE=k_cache.stride(0),
|
||||
FP8_MAX=448.0,
|
||||
USE_FP4=use_fp4_cache,
|
||||
MX_SCALE=not use_fp4_cache and mqa_logits_use_mx_scales(),
|
||||
SHUFFLE=shuffle,
|
||||
BLOCK_TILE_SIZE=_BLOCK_TILE_SIZE,
|
||||
HEAD_TILE_SIZE=_HEAD_TILE_SIZE,
|
||||
@@ -139,6 +143,7 @@ def _indexer_k_norm_rope_quant_store_kernel(
|
||||
KV_BLOCK_STRIDE: tl.constexpr,
|
||||
FP8_MAX: tl.constexpr,
|
||||
USE_FP4: tl.constexpr,
|
||||
MX_SCALE: tl.constexpr,
|
||||
SHUFFLE: tl.constexpr,
|
||||
BLOCK_TILE_SIZE: tl.constexpr,
|
||||
HEAD_TILE_SIZE: tl.constexpr,
|
||||
@@ -252,4 +257,10 @@ def _indexer_k_norm_rope_quant_store_kernel(
|
||||
tl.store(val_ptr + tiled, x_uint8)
|
||||
else:
|
||||
tl.store(val_ptr + block, x_uint8)
|
||||
tl.store(scale_ptr.to(tl.pointer_type(tl.float32)), tl.exp2(exponent))
|
||||
if MX_SCALE:
|
||||
tl.store(
|
||||
scale_ptr.to(tl.pointer_type(tl.int32)),
|
||||
pack_ue8m0_x4(tl.exp2(exponent)),
|
||||
)
|
||||
else:
|
||||
tl.store(scale_ptr.to(tl.pointer_type(tl.float32)), tl.exp2(exponent))
|
||||
|
||||
@@ -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),
|
||||
)
|
||||
|
||||
|
||||
@@ -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,
|
||||
|
||||
Reference in New Issue
Block a user