Performance: NVFP4 encoder slower than FP8 with TensorRT 11.4 on RTX PRO 6000 Blackwell, from copies and unfused quantize kernels around the FP4 GEMMs

未关闭
#4,880 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 6 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
8/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
cpp, python
领域
performance

调研方向

Start by reproducing the NVFP4 and FP8 engine builds with the trtexec commands in the report and comparing the per-layer profile. The work centres on the Myelin kernels named in the report (__myl_SlicResh, __myl_TranMul, __myl_Dyna_...) and the Q/K/V layout feeding the fused MHA kernel, which the payload does not map to source files. Done would mean the NVFP4 engine no longer pays for those extra copies and unfused quantize kernels, which is a compiler-level change.

由索引模型根据 Issue 内容生成。

描述

Module:Performance

Description

BAAI/bge-reranker-v2-m3 (XLM-R large, 24 layers) with NVFP4 W4A4 (ModelOpt NVFP4_DEFAULT_CFG, Q/K/V exported as one MatMul to avoid the batch fault in #4879) runs 45% slower than its FP8 engine at 128 pairs × 320 tokens and 24% slower at 128 × 640. Its FP4 GEMMs are not the problem: at 1,140–1,500 TFLOPS they are 1.7–2.0× faster than the FP8 GEMMs of the FP8 engine, about what CUTLASS reaches on this GPU outside TensorRT. The kernels that TensorRT adds around them cost more than the GEMMs save.

Per-layer profile (trtexec --dumpProfile --separateProfileRun), ms per inference at 128 × 640, grouped by the operations in the Myelin kernel names:

Kernel group FP8 engine NVFP4 engine
GEMMs 69.2 38.9
Q/K/V slicing, transposes and copies 0.9 38.3
FP4 dynamic quantize as its own kernel 0 18.0
residual add + LayerNorm (+ quantize) 8.6 18.0
GELU (+ quantize) 15.5 21.6
fused MHA 66.5 63.2
FP4 block-scale swizzle 0 2.1
total 160.8 200.1
  1. Q/K/V layout. In the FP16 and FP8 engines the Q/K/V GEMM writes [3, M, 1024] and _gemm_mha_v2 reads it directly, with 2 data-movement kernels in the whole engine. With the FP4 GEMM (cutlass3x_sm120_bstensorop_s16864gemm_block_scaled_ue4m3xe2m1_...), the GEMM writes [M, 3072], and five Myelin kernels per layer build the head-major Q, Kᵀ and V for the MHA kernel (__myl_SlicResh, __myl_TranMul, __myl_ReshMove, __myl_SlicReshTranReshMove, __myl_SlicReshTranMulReshMove). Together they take 1.6 ms per layer, about as long as all four GEMMs of the layer. With Q/K/V as separate MatMuls, the three FP4 GEMMs are not horizontally fused either (6 GEMMs and 11 data-movement kernels per layer).
  2. GELU and quantize. The FP4 dynamic quantize of the FFN-down input runs as its own kernel after GELU (__myl_Dyna_..., 0.61 ms per layer), reading back the FP16 GELU output. FP8 fuses GELU and the quantize into one kernel. #4735 reports the same separate __myl_Dyna kernel after a GEMM with a bias.
  3. LayerNorm and quantize. LayerNorm fused with the FP4 dynamic quantize (__myl_AddCastMeanSubMulMeanAddSqrtDivMulCast...Dyna_cuda_tile_...) takes 0.36 ms, twice the 0.17 ms of LayerNorm with the FP8 quantize.
  4. TensorRT 11.3.0.99 and 11.4.0.106 produce the same kernel mix and the same times (83.1 / 199.4 ms at 128 × 320 / 128 × 640).

Without the extra kernels and with the quantize fused like FP8, the NVFP4 engine would take about 127 ms at 128 × 640, 21% less than FP8, instead of 24% more.

Environment

TensorRT Version: 11.4.0.106 (also 11.3.0.99)

NVIDIA GPU: RTX PRO 6000 Blackwell Server Edition (SM120)

NVIDIA Driver Version: 595.91.07

CUDA Version: 13.4

CUDNN Version: not used

Operating System: Ubuntu 24.04 (GCP g4-standard-48)

Python Version (if applicable): 3.12

PyTorch Version (if applicable): 2.14.1 (export only)

Baremetal or Container (if so, version): nvcr.io/nvidia/tritonserver:26.09-py3

Relevant Files

Model link: BAAI/bge-reranker-v2-m3, quantized with ModelOpt NVFP4_DEFAULT_CFG (classifier head excluded) and exported to ONNX with get_onnx_bytes_and_metadata(..., weights_dtype="fp16"). The ONNX files (about 700 MB each) can be shared on request.

Steps To Reproduce

Commands or scripts:

trtexec --onnx=model.onnx --saveEngine=model.plan --minShapes=input_ids:1x128,attention_mask:1x128 \
  --optShapes=input_ids:32x640,attention_mask:32x640 --maxShapes=input_ids:128x768,attention_mask:128x768 \
  --profilingVerbosity=detailed --skipInference
trtexec --loadEngine=model.plan --shapes=input_ids:128x640,attention_mask:128x640 \
  --loadInputs=input_ids:ids.bin,attention_mask:mask.bin --warmUp=2000 --duration=10 --noDataTransfers \
  --useSpinWait --separateProfileRun --dumpProfile --exportProfile=profile.json --exportLayerInfo=layers.json

Have you tried the latest release?: Yes, 11.4.0.106 behaves the same.

Can this model run on other frameworks? Yes; the PyTorch fake-quant model and the engines agree in accuracy. This report is about speed only.

主要语言
C++
星标
13.5k
派生
2.4k
平均合并
6 天 21 小时
30 天内合并 PR
3

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

NVIDIA/TensorRT 的其他 Issue

查看 NVIDIA/TensorRT 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。