A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 03:15 UTC
[PERF] Remove TRTLLM Gen attn kernel limitation `max_seq_len <=131072` (#28755)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Vadim Gimpelson committed
Nov 15, 2025 at 14:13 UTC
173b356abff3e2e547fc44c60361f3b0adc41aaf
2 files changed
+2
-19
vllm/config/vllm.py
-15
@@ -483,21 +483,6 @@ class VllmConfig:
483
"Overriding cudagraph_mode to PIECEWISE."
484
)
485
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
486
- elif (
487
- current_platform.is_cuda()
488
- and current_platform.is_device_capability(100)
489
- and self.model_config.max_model_len > 131072
490
- and not self.model_config.use_mla
491
- ):
492
- # Refer to vllm/utils/flashinfer.py::use_trtllm_attention()
493
- logger.warning_once(
494
- "NVIDIA Blackwell TRTLLM attention cannot support "
495
- "max_model_len >= 131072 (found "
496
- f"{self.model_config.max_model_len}), causing dynamic "
497
- "dispatching that breaks full cudagraphs. "
498
- "Overriding cudagraph_mode to PIECEWISE."
499
- )
500
- self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
486
487
# disable cudagraph when enforce eager execution
488
if self.model_config is not None and self.model_config.enforce_eager:
vllm/utils/flashinfer.py
+2
-4
@@ -319,14 +319,12 @@ def use_trtllm_attention(
319
# Environment variable not set - use auto-detection
320
if is_prefill:
321
# Prefill auto-detection
322
- use_trtllm = max_seq_len <= 131072 and kv_cache_dtype == "auto"
322
+ use_trtllm = kv_cache_dtype == "auto"
323
if use_trtllm:
324
logger.warning_once("Using TRTLLM prefill attention (auto-detected).")
325
else:
326
# Decode auto-detection
327
- use_trtllm = (
328
- num_tokens <= 256 and max_seq_len <= 131072 and kv_cache_dtype == "auto"
329
- )
327
+ use_trtllm = num_tokens <= 256 and kv_cache_dtype == "auto"
328
if use_trtllm:
329
logger.warning_once("Using TRTLLM decode attention (auto-detected).")
330
return use_trtllm