Harden DeepSeek V3.2 fused kernel grids (#52381)

Signed-off-by: yimdev <5779256+yimdev@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
yimdev
2026-08-18 01:43:12 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent e8ad2855e7
commit eab1cff5b0
2 changed files with 79 additions and 6 deletions
+6 -6
View File
@@ -152,8 +152,8 @@ def _fused_norm_rope_kernel(
INDEX_ROPE_INTERLEAVE: tl.constexpr,
USE_PDL: tl.constexpr,
):
pid = tl.program_id(0)
tok_idx = tl.program_id(1)
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()
@@ -479,7 +479,7 @@ def fused_norm_rope(
if q_c_out is None:
q_c_out = torch.empty_like(q_c)
use_pdl = current_platform.is_arch_support_pdl()
_fused_norm_rope_kernel[(4, num_tokens)](
_fused_norm_rope_kernel[(num_tokens, 4)](
positions,
# Q RMS norm
q_c,
@@ -594,8 +594,8 @@ def _fused_q_kernel(
QUANTIZE_MQA: tl.constexpr,
USE_PDL: tl.constexpr,
):
pid = tl.program_id(0)
tok_idx = tl.program_id(1)
tok_idx = tl.program_id(0).to(tl.int64)
pid = tl.program_id(1)
head_idx = tl.program_id(2)
if USE_PDL:
tl.extra.cuda.gdc_wait()
@@ -869,7 +869,7 @@ def fused_q(
return index_q_fp8, index_weights_out, mqa_q
use_pdl = current_platform.is_arch_support_pdl()
_fused_q_kernel[(3, num_tokens, grid_heads)](
_fused_q_kernel[(num_tokens, 3, grid_heads)](
positions,
q_pe,
q_pe.stride(0),