def qwen4_exp_qsa_triton_warmup(worker: "Worker") -> None:
"""Warm every reachable QSA specialization: indexer decode-query-length
profiles plus the sparse attention split-K/merge configs."""
qsa_module = sys.modules.get("vllm.models.qwen4_exp.nvidia.qsa_indexer")
attn_module = sys.modules.get("vllm.models.qwen4_exp.nvidia.qsa")
if qsa_module is None or attn_module is None:
return
indexer = None
owner = None
for layer in worker.get_model().modules():
if indexer is None and isinstance(layer, qsa_module.QSAIndexer):
indexer = layer
elif owner is None and isinstance(layer, attn_module.Qwen4ExpQSAAttention):
owner = layer
if indexer is None or owner is None:
return
runner = worker.model_runner
def block_table_for(prefix: str) -> torch.Tensor:
group_id = next(
i
for i, group in enumerate(runner.kv_cache_config.kv_cache_groups)
if prefix in group.layer_names
)
return runner.block_tables.input_block_tables[group_id]
from vllm.models.qwen4_exp.nvidia.ops.qsa_indexer import (
qsa_prefill_logits_width,
use_gvr2_topk,
warmup_gvr2_topk,
warmup_qsa_mqa_paged_decode,
)
k_cache = indexer.compressed_key_cache.kv_cache
assert k_cache.numel()
profiles = warmup_qsa_mqa_paged_decode(
k_cache,
block_table_for(indexer.compressed_key_cache.prefix),
num_heads=indexer.index_n_heads,
head_dim=indexer.index_head_dim,
max_decode_query_len=runner.decode_query_len,
max_num_reqs=runner.max_num_reqs,
max_num_batched_tokens=runner.max_num_tokens,
)
logger.info("Warmed up Qwen4Exp QSA decode kernels: %s.", profiles)
# Pre-compile gvr_2 so neither CUDA-graph capture nor serving compiles it.
block_topk = indexer.token_topk // indexer.compress_ratio
decode_width = (
block_table_for(indexer.compressed_key_cache.prefix).shape[1] * k_cache.shape[1]
)
prefill_width = qsa_prefill_logits_width(
worker.vllm_config.model_config.max_model_len,
indexer.compress_ratio,
decode_width,
)
# prefill_width is float4-aligned, so this only checks platform and k.
if use_gvr2_topk(indexer.topk_backend, block_topk, prefill_width):
warmup_gvr2_topk(
indexer.topk_backend,
block_topk,
decode_width=decode_width,
max_decode_rows=runner.max_num_reqs * runner.decode_query_len,
prefill_width=prefill_width,
)
logger.info(
"Warmed up Qwen4Exp QSA gvr_2 top-k (k=%d, decode width %d, "
"prefill width %d).",
block_topk,
decode_width,
prefill_width,
)
from vllm.models.qwen4_exp.nvidia.ops.qsa import (
warmup_qsa_sparse_paged_attention,
)
kv_cache = owner.kv_cache
assert kv_cache.numel()
attention_profiles = warmup_qsa_sparse_paged_attention(
kv_cache,
block_table_for(owner.layer_name),
num_query_heads=owner.num_heads,
selection_width=indexer.output_width,
)
logger.info(
"Warmed up Qwen4Exp QSA sparse attention kernels: %s.",
attention_profiles,
)