[ROCm][Perf] Enqueue the default chain first in CSA/HCA at small batch sizes for DeepSeek-V4 (#61378)

Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen
2026-10-14 10:57:26 +08:00
committed by GitHub
parent 2308b6f00b
commit 2521446ac8
+14
View File
@@ -162,6 +162,16 @@ def apply_pre_quantized_block_scaled_mm(
_SPARSE_PREFILL_TOPK_ALIGNMENT = 128
def _default_first_token_threshold() -> int:
"""Max tokens at which the CSA/HCA fork enqueues the default chain first.
At small batches starting the long default chain first shortens the step
tail; larger batches saturate the device and aux-first is faster. Only
tuned on gfx950 (24 tokens = conc 8 with MTP=2); 0 disables elsewhere.
"""
return 24 if _ON_GFX950 else 0
@triton.jit
def _combine_topk_swa_indices_kernel(
combined_indices_ptr,
@@ -807,6 +817,8 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
indexer = self.indexer
compressor = self.compressor
assert compressor is not None
# Static per captured graph: each capture size bakes its own order.
default_first = hidden_states.shape[0] <= _default_first_token_threshold()
def default_chain():
qr_kv = self._fused_wqa_wkv_gemm(hidden_states)
@@ -838,6 +850,7 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
[self.ln_events[1]],
aux_streams[:1],
enable=True,
default_first=default_first,
)
self._sparse_indexer_and_attn(
hidden_states, None, None, None, q, kv_out, positions, o_padded
@@ -864,6 +877,7 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
self.ln_events[1:3],
aux_streams[:2],
enable=True,
default_first=default_first,
)
indexer_weights_out, _ = indexer.weights_proj(hidden_states)