[ROCm][Model][Bugfix] Enable GLM-5.2-MXFP4 on the deepseek_v32 path and fix sparse attention correctness (#51915)

Signed-off-by: Jack Hu <Jack.Hu@amd.com>
Signed-off-by: Jack Hu <jack.hu@amd.com>
Signed-off-by: Douglas Lehr <Doug.Lehr@amd.com>
Co-authored-by: James E T Smith <jamesETsmith@users.noreply.github.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
Co-authored-by: Douglas Lehr <Doug.Lehr@amd.com>
This commit is contained in:
Jack Hu
2026-09-22 18:08:51 -05:00
committed by GitHub
co-authored by James E T Smith Douglas Lehr Douglas Lehr
parent 9646f53064
commit 1c0eee919d
8 changed files with 318 additions and 211 deletions
+8 -4
View File
@@ -7,9 +7,9 @@ MLA + a "lightning indexer" that selects the top-k tokens for a sparse MLA
attend. The same model code serves any DSA checkpoint, including GLM-5.2
(``glm_moe_dsa``), which reuses this architecture.
The CUDA implementation selects capability-specific kernels internally and
falls back when an optimization is unavailable. Other platforms use the
generic implementation by default.
The CUDA and ROCm implementations select capability-specific kernels
internally and fall back when an optimization is unavailable. Other platforms
use the generic implementation by default.
"""
from vllm.platforms import current_platform
@@ -20,8 +20,12 @@ if current_platform.is_cuda():
from .nvidia.model import DeepseekV32ForCausalLM
from .nvidia.model import DeepseekV32ForCausalLM as GlmMoeDsaForCausalLM
from .nvidia.mtp import DeepseekV32MTP
elif current_platform.is_rocm():
from .amd.model import DeepseekV32ForCausalLM
from .amd.model import DeepseekV32ForCausalLM as GlmMoeDsaForCausalLM
from .amd.mtp import DeepseekV32MTP
else:
# ROCm, XPU, and CPU keep the generic implementation.
# XPU and CPU keep the generic implementation.
from vllm.model_executor.models.deepseek_mtp import DeepSeekMTP as DeepseekV32MTP
from vllm.model_executor.models.deepseek_v2 import (
DeepseekV3ForCausalLM as DeepseekV32ForCausalLM,
+16 -25
View File
@@ -3,12 +3,14 @@
import torch
from vllm._aiter_ops import rocm_aiter_ops
from vllm.compilation.breakable_cudagraph import eager_break_during_capture
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer
from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
from vllm.models.deepseek_v32.attention import DeepseekV32Attention, DeepseekV32Indexer
from vllm.models.deepseek_v32.common.kernels import fused_norm_rope, fused_q
from vllm.platforms import current_platform
from vllm.utils.torch_utils import is_quantized_kv_cache
from vllm.v1.attention.backends.mla.indexer import DeepseekV32IndexerBackend
from vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse import (
@@ -16,21 +18,15 @@ from vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse import (
)
class DeepseekV32MLASparseBackend(ROCMAiterMLASparseBackend):
@staticmethod
def get_supported_kernel_block_sizes() -> list:
return [16, 32]
class DeepseekV32ROCmIndexerBackend(DeepseekV32IndexerBackend):
@staticmethod
def get_supported_kernel_block_sizes() -> list:
return [16, 32]
class DeepseekV32ROCmIndexerCache(DeepseekV32IndexerCache):
def get_attn_backend(self):
return DeepseekV32ROCmIndexerBackend
return DeepseekV32IndexerBackend
@property
def uses_shuffled_layout(self) -> bool:
# aiter's gather/insert pair shuffles the cache above block size 1:
# [n_blocks, blk/16, head_dim/16, 16, 16] instead of [n_blocks, blk, head_dim].
return self.kv_cache.ndim == 3 and self.kv_cache.shape[1] != 1
class DeepseekV32ROCmIndexer(DeepseekV32Indexer):
@@ -46,7 +42,7 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
config,
prefix,
topk_indices_buffer,
attn_backend=DeepseekV32MLASparseBackend,
attn_backend=ROCMAiterMLASparseBackend,
)
self.indexer_op: SparseAttnIndexer | None = None
@@ -100,14 +96,10 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
q_nope = q_nope.transpose(0, 1) # (N, tokens, P)
if self.is_aiter_triton_fp4_bmm_enabled:
from aiter.ops.triton.batched_gemm_a16wfp4 import batched_gemm_a16wfp4
ql_nope = batched_gemm_a16wfp4(
ql_nope = rocm_aiter_ops.batched_gemm_a16wfp4(
q_nope, self.W_K, self.W_K_scale, transpose_bm=True, prequant=True
)
elif self.is_aiter_triton_fp8_bmm_enabled:
from vllm._aiter_ops import rocm_aiter_ops
ql_nope = rocm_aiter_ops.triton_fp8_bmm(
q_nope, self.W_K, self.W_K_scale, group_size=128, transpose_bm=True
)
@@ -152,14 +144,10 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
)
if self.is_aiter_triton_fp4_bmm_enabled:
from aiter.ops.triton.batched_gemm_a16wfp4 import batched_gemm_a16wfp4
batched_gemm_a16wfp4(
rocm_aiter_ops.batched_gemm_a16wfp4(
x, self.W_V, self.W_V_scale, out_view, transpose_bm=True, prequant=True
)
elif self.is_aiter_triton_fp8_bmm_enabled:
from vllm._aiter_ops import rocm_aiter_ops
rocm_aiter_ops.triton_fp8_bmm(
x,
self.W_V,
@@ -202,6 +190,7 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
indexer_k_norm_eps = self.indexer.k_norm.eps
indexer_k_rope_cos_sin_cache = self.indexer_rope_emb.cos_sin_cache
indexer_k_cache = self.indexer.k_cache.kv_cache
indexer_cache_shuffled = self.indexer.k_cache.uses_shuffled_layout
indexer_softmax_scale = self.indexer.softmax_scale
indexer_n_head_scale = self.indexer.n_head**-0.5
else:
@@ -211,6 +200,7 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
indexer_k_norm_eps = 1e-6
indexer_k_rope_cos_sin_cache = None
indexer_k_cache = None
indexer_cache_shuffled = False
indexer_softmax_scale = 0.0
indexer_n_head_scale = 0.0
@@ -241,6 +231,7 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
self.topk_indices_buffer,
slot_mapping=mla_slot,
indexer_k_cache=indexer_k_cache,
indexer_cache_shuffled=indexer_cache_shuffled,
mla_kv_cache=mla_kv_cache,
mla_kv_cache_dtype=self.kv_cache_dtype,
mla_k_scale=mla_k_scale,
@@ -282,7 +273,7 @@ class DeepseekV32MLAAttention(DeepseekV32Attention):
num_actual = attn_metadata.num_actual_tokens # type: ignore[attr-defined]
kv_cache = self.kv_cache
if self._fp8_kv_needs_view:
kv_cache = kv_cache.view(torch.float8_e4m3fn)
kv_cache = kv_cache.view(current_platform.fp8_dtype())
q_for_attn = self._build_q_for_attn(ql_nope, mqa_q, num_actual)
attn_out, _ = self.impl.forward_mqa( # type: ignore[attr-defined]
+4 -1
View File
@@ -327,6 +327,7 @@ class DeepseekV32Attention(MLAAttention):
indexer_k_norm_bias = self.indexer.k_norm.bias
indexer_k_norm_eps = self.indexer.k_norm.eps
indexer_k_rope_cos_sin_cache = self.indexer_rope_emb.cos_sin_cache
indexer_cache_shuffled = self.indexer.k_cache.uses_shuffled_layout
indexer_k_cache = None if self.use_pcp else self.indexer.k_cache.kv_cache
index_k_out = torch.empty_like(index_k) if self.use_pcp else None
indexer_softmax_scale = self.indexer.softmax_scale
@@ -338,6 +339,7 @@ class DeepseekV32Attention(MLAAttention):
indexer_k_norm_eps = 1e-6
indexer_k_rope_cos_sin_cache = None
indexer_k_cache = None
indexer_cache_shuffled = False
index_k_out = None
indexer_softmax_scale = 0.0
indexer_n_head_scale = 0.0
@@ -377,6 +379,7 @@ class DeepseekV32Attention(MLAAttention):
slot_mapping=mla_slot,
indexer_slot_mapping=indexer_slot,
indexer_k_cache=indexer_k_cache,
indexer_cache_shuffled=indexer_cache_shuffled,
mla_kv_cache=mla_kv_cache,
mla_kv_cache_dtype=self.kv_cache_dtype,
mla_k_scale=mla_k_scale,
@@ -544,7 +547,7 @@ class DeepseekV32Attention(MLAAttention):
return
if self._fp8_kv_needs_view:
kv_cache = kv_cache.view(torch.float8_e4m3fn)
kv_cache = kv_cache.view(current_platform.fp8_dtype())
if self._fp8_query:
# FlashInfer sparse: single packed fp8 query.
mqa_q_arg: torch.Tensor | tuple[torch.Tensor, torch.Tensor] = mqa_q[
+97 -27
View File
@@ -12,6 +12,18 @@ from vllm.utils.torch_utils import is_quantized_kv_cache
# has_indexer=False path so the indexer args don't allocate every call.
_DUMMY_CACHE: dict[tuple, torch.Tensor] = {}
# Tile shape of the indexer-K cache's shuffled layout, used when the cache reports
# uses_shuffled_layout. The block tile is a token count; the head tile is a byte
# count, which the kernel converts to cache elements.
_INDEXER_CACHE_BLOCK_TILE = 16
_INDEXER_CACHE_HEAD_TILE_BYTES = 16
# FNUZ (e4m3fnuz) on gfx942, OCP (e4m3fn) elsewhere. 224.0 keeps the ue8m0
# power-of-two scale inside the FNUZ range after its ceil.
_FP8_DTYPE = current_platform.fp8_dtype()
_USE_FNUZ = torch.float8_e4m3fnuz == _FP8_DTYPE
_FP8_MAX = 224.0 if _USE_FNUZ else 448.0
def _dummy(shape: tuple, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
key = (shape, dtype, device)
@@ -47,16 +59,17 @@ def _get_cos_sin(
@triton.jit
def _fp8_ue8m0_quantize(vals):
def _fp8_ue8m0_quantize(vals, FP8_MAX: tl.constexpr, USE_FNUZ: tl.constexpr):
"""Quantize float32 values to FP8 E4M3 with a ue8m0 (power-of-2) scale.
Returns (fp8_vals, scale) so the caller can store them or reuse the scale.
"""
vals = vals.to(tl.float32)
amax = tl.max(tl.abs(vals))
scale = tl.div_rn(tl.maximum(amax, 1e-4), 448.0)
scale = tl.div_rn(tl.maximum(amax, 1e-4), FP8_MAX)
scale = tl.math.exp2(tl.math.ceil(tl.math.log2(scale)))
fp8_vals = tl.div_rn(vals, scale).to(tl.float8e4nv)
fp8_dtype = tl.float8e4b8 if USE_FNUZ else tl.float8e4nv
fp8_vals = tl.div_rn(vals, scale).to(fp8_dtype)
return fp8_vals, scale
@@ -71,18 +84,32 @@ def _fp8_quant_and_cache_write(
cache_block_stride,
offsets,
HEAD_DIM: tl.constexpr,
SHUFFLE: tl.constexpr,
BLOCK_TILE: tl.constexpr,
HEAD_TILE: tl.constexpr,
FP8_MAX: tl.constexpr,
USE_FNUZ: tl.constexpr,
):
k_fp8, scale = _fp8_ue8m0_quantize(vals)
k_fp8, scale = _fp8_ue8m0_quantize(vals, FP8_MAX, USE_FNUZ)
block_idx = slot_idx // cache_block_size
block_offset = slot_idx % cache_block_size
block_start = block_idx * cache_block_stride
tl.store(
kv_cache_ptr + block_start + block_offset * HEAD_DIM + offsets,
k_fp8,
mask=mask,
)
# Shuffled layout: [blk/BLOCK_TILE, HEAD_DIM/HEAD_TILE, BLOCK_TILE, HEAD_TILE],
# so one contiguous run holds BLOCK_TILE tokens x HEAD_TILE bytes and the
# reader's coalesced loads land on the bytes we wrote.
if SHUFFLE:
value_off = (
block_offset // BLOCK_TILE * BLOCK_TILE * HEAD_DIM
+ block_offset % BLOCK_TILE * HEAD_TILE
+ offsets // HEAD_TILE * BLOCK_TILE * HEAD_TILE
+ offsets % HEAD_TILE
)
else:
value_off = block_offset * HEAD_DIM + offsets
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
tl.store(kv_cache_scale_ptr + scale_byte_off // 4, scale)
@@ -137,6 +164,9 @@ def _fused_norm_rope_kernel(
indexer_cache_scale_ptr,
indexer_cache_block_size,
indexer_cache_stride,
INDEXER_CACHE_SHUFFLE: tl.constexpr,
INDEXER_CACHE_BLOCK_TILE: tl.constexpr,
INDEXER_CACHE_HEAD_TILE: tl.constexpr,
# MLA KV cache (concat kv_c_normed + k_pe_roped, uses slot_mapping_ptr)
mla_cache_ptr,
mla_cache_block_stride,
@@ -161,12 +191,15 @@ def _fused_norm_rope_kernel(
HAS_INDEXER: tl.constexpr,
INDEX_ROPE_INTERLEAVE: tl.constexpr,
USE_PDL: tl.constexpr,
FP8_MAX: tl.constexpr,
USE_FNUZ: tl.constexpr,
):
tok_idx = tl.program_id(0).to(tl.int64)
pid = tl.program_id(1)
if USE_PDL:
tl.extra.cuda.gdc_wait()
tl.extra.cuda.gdc_launch_dependents()
fp8_dtype = tl.float8e4b8 if USE_FNUZ else tl.float8e4nv
if pid == 3:
if not HAS_INDEXER:
# Shared layer: reuse the previous indexer layer's top-k; do not
@@ -327,9 +360,11 @@ def _fused_norm_rope_kernel(
)
kv_2d = tl.reshape(kv_c, (MLA_NUM_TILES, MLA_TILE_DIM))
tile_amax = tl.max(tl.abs(kv_2d), axis=1, keep_dims=True)
tile_scale = tl.maximum(tile_amax * (1.0 / 448.0), 1e-4)
# scale = amax / FP8_MAX (fp8 e4m3 max), matching the reference
# concat_and_cache_ds_mla kernel; floored to FLT_MIN.
tile_scale = tl.maximum(tile_amax * (1.0 / FP8_MAX), 1e-4)
tile_scale = tl.math.exp2(tl.math.ceil(tl.math.log2(tile_scale)))
kv_c_fp8 = tl.reshape((kv_2d / tile_scale).to(tl.float8e4nv), (KV_DIM,))
kv_c_fp8 = tl.reshape((kv_2d / tile_scale).to(fp8_dtype), (KV_DIM,))
tl.store(mla_cache_ptr + byte_base + kv_block, kv_c_fp8)
tile_off = tl.arange(0, MLA_NUM_TILES)
tl.store(
@@ -349,16 +384,16 @@ def _fused_norm_rope_kernel(
# kv_c_normed (KV_DIM elements)
if MLA_CACHE_FP8:
scale = tl.load(mla_cache_scale_ptr)
kv_c_fp8 = (kv_c.to(tl.float32) / scale).to(tl.float8e4nv)
kv_c_fp8 = (kv_c.to(tl.float32) / scale).to(fp8_dtype)
tl.store(dst + kv_block, kv_c_fp8)
else:
tl.store(dst + kv_block, kv_c)
# k_pe_roped (from registers, interleaved layout)
if MLA_CACHE_FP8:
tl.store(dst + KV_DIM + dim_off * 2, (r1 / scale).to(tl.float8e4nv))
tl.store(dst + KV_DIM + dim_off * 2, (r1 / scale).to(fp8_dtype))
tl.store(
dst + KV_DIM + dim_off * 2 + 1,
(r2 / scale).to(tl.float8e4nv),
(r2 / scale).to(fp8_dtype),
)
else:
tl.store(dst + KV_DIM + dim_off * 2, r1)
@@ -463,6 +498,11 @@ def _fused_norm_rope_kernel(
indexer_cache_stride,
index_k_block,
INDEX_K_DIM,
INDEXER_CACHE_SHUFFLE,
INDEXER_CACHE_BLOCK_TILE,
INDEXER_CACHE_HEAD_TILE,
FP8_MAX,
USE_FNUZ,
)
@@ -486,6 +526,7 @@ def fused_norm_rope(
slot_mapping: torch.Tensor | None = None,
indexer_slot_mapping: torch.Tensor | None = None,
indexer_k_cache: torch.Tensor | None = None,
indexer_cache_shuffled: bool = False,
mla_kv_cache: torch.Tensor | None = None,
mla_kv_cache_dtype: str = "auto",
mla_k_scale: torch.Tensor | None = None,
@@ -532,13 +573,28 @@ def fused_norm_rope(
indexer_slot_mapping = slot_mapping
idx_cache_scale_view = indexer_k_cache.view(torch.uint8).view(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
# layout; see DeepseekV32IndexerCache.uses_shuffled_layout.
idx_cache_shuffle = indexer_cache_shuffled
if idx_cache_shuffle:
assert idx_cache_block_size % _INDEXER_CACHE_BLOCK_TILE == 0, (
f"indexer K cache block size {idx_cache_block_size} must be a "
f"multiple of {_INDEXER_CACHE_BLOCK_TILE} for the shuffled layout"
)
idx_cache_stride = indexer_k_cache.stride(0)
if indexer_k_cache.dtype == torch.uint8:
indexer_k_cache = indexer_k_cache.view(torch.float8_e4m3fn)
indexer_k_cache = indexer_k_cache.view(_FP8_DTYPE)
# The head tile is a byte count; the kernel indexes the cache in elements.
idx_cache_head_tile = (
_INDEXER_CACHE_HEAD_TILE_BYTES // indexer_k_cache.element_size()
)
else:
idx_cache_scale_view = None
idx_cache_block_size = 1
idx_cache_stride = 0
idx_cache_shuffle = False
idx_cache_head_tile = _INDEXER_CACHE_HEAD_TILE_BYTES
# --- MLA KV cache setup ---
mla_cache_nvfp4 = mla_kv_cache_dtype == "nvfp4_ds_mla"
@@ -561,12 +617,12 @@ def fused_norm_rope(
mla_entry_stride = u8_cache.stride(1)
mla_ds_scale_view = u8_cache.view(torch.float32)
mla_ds_rope_view = u8_cache.view(torch.bfloat16)
mla_kv_cache = u8_cache.view(torch.float8_e4m3fn)
mla_kv_cache = u8_cache.view(_FP8_DTYPE)
else:
mla_block_stride = mla_kv_cache.stride(0)
mla_entry_stride = mla_kv_cache.stride(1)
if mla_cache_fp8 and mla_kv_cache.dtype == torch.uint8:
mla_kv_cache = mla_kv_cache.view(torch.float8_e4m3fn)
mla_kv_cache = mla_kv_cache.view(_FP8_DTYPE)
if mla_k_scale is None:
mla_k_scale = torch.ones(1, dtype=torch.float32, device=device)
else:
@@ -592,6 +648,9 @@ def fused_norm_rope(
assert index_k_out.shape == index_k.shape
index_k_out_stride = index_k_out.stride(0)
use_pdl = current_platform.is_arch_support_pdl()
pdl_kwargs = (
{"USE_PDL": use_pdl, "launch_pdl": True} if use_pdl else {"USE_PDL": use_pdl}
)
_fused_norm_rope_kernel[(num_tokens, 4)](
positions,
# Q RMS norm
@@ -639,6 +698,9 @@ def fused_norm_rope(
idx_cache_scale_view,
idx_cache_block_size,
idx_cache_stride,
idx_cache_shuffle,
_INDEXER_CACHE_BLOCK_TILE,
idx_cache_head_tile,
# MLA KV cache (uses same slot_mapping)
mla_kv_cache,
mla_block_stride,
@@ -659,8 +721,9 @@ def fused_norm_rope(
TOPK_BLOCK_SIZE=1024,
HAS_INDEXER=has_indexer,
INDEX_ROPE_INTERLEAVE=index_rope_interleave,
USE_PDL=use_pdl,
launch_pdl=use_pdl,
FP8_MAX=_FP8_MAX,
USE_FNUZ=_USE_FNUZ,
**pdl_kwargs,
)
return q_c_out
@@ -715,6 +778,8 @@ def _fused_q_kernel(
INDEX_ROPE_INTERLEAVE: tl.constexpr,
QUANTIZE_MQA: tl.constexpr,
USE_PDL: tl.constexpr,
FP8_MAX: tl.constexpr,
USE_FNUZ: tl.constexpr,
):
tok_idx = tl.program_id(0).to(tl.int64)
pid = tl.program_id(1)
@@ -722,6 +787,7 @@ def _fused_q_kernel(
if USE_PDL:
tl.extra.cuda.gdc_wait()
tl.extra.cuda.gdc_launch_dependents()
fp8_dtype = tl.float8e4b8 if USE_FNUZ else tl.float8e4nv
if pid == 2:
# ql_nope quantize + pack into the front of mqa_q_fp8. On the bf16
@@ -744,7 +810,7 @@ def _fused_q_kernel(
+ ql_nope_off,
mask=ql_nope_mask,
).to(tl.float32)
ql_nope_fp8 = (ql_nope / scale).to(tl.float8e4nv)
ql_nope_fp8 = (ql_nope / scale).to(fp8_dtype)
tl.store(
mqa_q_fp8_ptr
+ tok_idx * mqa_q_fp8_stride0
@@ -794,7 +860,7 @@ def _fused_q_kernel(
+ q_head_idx * mqa_q_fp8_stride1
+ QL_NOPE_DIM
+ rot_off * 2,
(r1 / scale).to(tl.float8e4nv),
(r1 / scale).to(fp8_dtype),
)
tl.store(
mqa_q_fp8_ptr
@@ -803,7 +869,7 @@ def _fused_q_kernel(
+ QL_NOPE_DIM
+ rot_off * 2
+ 1,
(r2 / scale).to(tl.float8e4nv),
(r2 / scale).to(fp8_dtype),
)
else:
# bf16 query: write the RoPE'd q_pe unquantized.
@@ -861,7 +927,7 @@ def _fused_q_kernel(
index_q = tl.where(in_rope, roped, index_q)
# Index Q Quantize (from registers)
index_q_fp8, index_q_scale = _fp8_ue8m0_quantize(index_q)
index_q_fp8, index_q_scale = _fp8_ue8m0_quantize(index_q, FP8_MAX, USE_FNUZ)
tl.store(
index_q_fp8_ptr
+ tok_idx * index_q_fp8_stride0
@@ -955,7 +1021,7 @@ def fused_q(
q_pe.shape[0],
q_pe.shape[1],
ql_nope.shape[2] + q_pe.shape[2],
dtype=torch.float8_e4m3fn,
dtype=_FP8_DTYPE,
device=q_pe.device,
)
# Placeholder; pid 0 packs q_pe into mqa_q_fp8 instead.
@@ -967,7 +1033,7 @@ def fused_q(
mqa_q_fp8 = q_pe_out # unused placeholder for the fp8 pack pointer
mqa_q = q_pe_out
index_q_fp8 = torch.empty_like(index_q, dtype=torch.float8_e4m3fn)
index_q_fp8 = torch.empty_like(index_q, dtype=_FP8_DTYPE)
index_weights_out = torch.empty_like(index_weights, dtype=torch.float32)
if cutedsl_kernel is not None:
cutedsl_kernel(
@@ -990,6 +1056,9 @@ def fused_q(
return index_q_fp8, index_weights_out, mqa_q
use_pdl = current_platform.is_arch_support_pdl()
pdl_kwargs = (
{"USE_PDL": use_pdl, "launch_pdl": True} if use_pdl else {"USE_PDL": use_pdl}
)
_fused_q_kernel[(num_tokens, 3, grid_heads)](
positions,
q_pe,
@@ -1031,8 +1100,9 @@ def fused_q(
HAS_INDEXER=has_indexer,
INDEX_ROPE_INTERLEAVE=index_rope_interleave,
QUANTIZE_MQA=quantize_mqa,
USE_PDL=use_pdl,
launch_pdl=use_pdl,
FP8_MAX=_FP8_MAX,
USE_FNUZ=_USE_FNUZ,
**pdl_kwargs,
# num_warps=1 is optimal here: each program is a single 128-element
# rope+quant, so the kernel is program-count/occupancy bound, not
# per-program compute bound (swept 1/2/4/8 — 1 wins or ties everywhere).