[Perf][3/N] Expand Triton kernel warmup coverage, Qwen (#47546)
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Roberto L. Castro committed
Jul 7, 2026 at 19:06 UTC
c3284c31f52c005bde02cf7899959c2539b01d2f
1 file changed
+3
-6
vllm/model_executor/warmup/qwen_triton_warmup.py
+3
-6
@@ -206,8 +206,6 @@ def _warm_zero_kv_blocks_kernel(
206
N_SEGS=config.n_segs,
207
PAGE_SIZE_EL=config.page_size_el,
208
BLOCK_SIZE=config.block_size,
209
- num_warps=4,
210
- num_stages=3,
209
)
210
211
@@ -372,11 +370,10 @@ def qwen_triton_warmup(
370
logger.info("Warming up Qwen Triton kernels for model_type=%s.", model_type)
371
372
zero_config = _zero_kv_warmup_config(runner)
375
- if _warm_zero_kv_blocks_with_runner_zeroer(runner):
376
- pass
377
- elif zero_config is not None:
373
+ warmed_zeroer = _warm_zero_kv_blocks_with_runner_zeroer(runner)
374
+ if zero_config is not None:
375
_warm_zero_kv_blocks_kernel(device, zero_config)
379
- else:
376
+ elif not warmed_zeroer:
377
logger.info("Skipping Qwen zero-kv warmup: no KVBlockZeroer metadata.")
378
379
_warm_compute_slot_mapping_kernel(device)