[Deprecation] Remove dead use_prefill_decode_attention flag (#52557)
Signed-off-by: Brian Woods <me@brian.biz> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Brian Woods committed
Aug 22, 2026 at 15:38 UTC
bbe8b23e1a2b32a96240b27f63255170d09ef144
3 files changed
-10
tests/engine/test_arg_utils.py
-5
@@ -383,8 +383,6 @@ def test_attention_config():
383
"FLASH_ATTN",
384
"--attention-config.flash_attn_version",
385
"3",
386
- "--attention-config.use_prefill_decode_attention",
387
- "true",
386
"--attention-config.flash_attn_max_num_splits_for_cuda_graph",
387
"16",
388
"--attention-config.use_trtllm_attention",
@@ -398,7 +396,6 @@ def test_attention_config():
396
assert engine_args.attention_config.backend is not None
397
assert engine_args.attention_config.backend.name == "FLASH_ATTN"
398
assert engine_args.attention_config.flash_attn_version == 3
401
- assert engine_args.attention_config.use_prefill_decode_attention is True
399
assert engine_args.attention_config.flash_attn_max_num_splits_for_cuda_graph == 16
400
assert engine_args.attention_config.use_trtllm_attention is True
401
assert engine_args.attention_config.disable_flashinfer_q_quantization is True
@@ -408,7 +405,6 @@ def test_attention_config():
405
[
406
"--attention-config="
407
'{"backend": "FLASHINFER", "flash_attn_version": 2, '
411
- '"use_prefill_decode_attention": false, '
408
'"flash_attn_max_num_splits_for_cuda_graph": 8, '
409
'"use_trtllm_attention": false, '
410
'"disable_flashinfer_q_quantization": false}',
@@ -419,7 +415,6 @@ def test_attention_config():
415
assert engine_args.attention_config.backend is not None
416
assert engine_args.attention_config.backend.name == "FLASHINFER"
417
assert engine_args.attention_config.flash_attn_version == 2
422
- assert engine_args.attention_config.use_prefill_decode_attention is False
418
assert engine_args.attention_config.flash_attn_max_num_splits_for_cuda_graph == 8
419
assert engine_args.attention_config.use_trtllm_attention is False
420
assert engine_args.attention_config.disable_flashinfer_q_quantization is False
tests/v1/attention/utils.py
-1
@@ -387,7 +387,6 @@ full_cg_backend_configs = {
387
name="RocmAttn",
388
attention_config={
389
"backend": "ROCM_ATTN",
390
- "use_prefill_decode_attention": True,
390
},
391
comp_config={
392
"cudagraph_mode": "FULL",
vllm/config/attention.py
-4
@@ -42,10 +42,6 @@ class AttentionConfig:
42
"""Force vllm to use a specific flash-attention version (2, 3, or 4).
43
Only valid when using the flash-attention backend."""
44
45
- use_prefill_decode_attention: bool = False
46
- """Use separate prefill and decode kernels for attention instead of
47
- the unified triton kernel."""
48
-
45
flash_attn_max_num_splits_for_cuda_graph: int = 32
46
"""Flash Attention max number splits for cuda graph decode."""
47