GitLab

Bug description:
With torch 2.4.0.dev20240603+cu121,
cutlass_fp8_supported outputs False, and the (capability, version) before the comparison is (90, 11111111112)

This PR fixes the support check for FP8 CUTLASS ( cutlass_fp8_supported) which was introduced in https://github.com/vllm-project/vllm/pull/5183.

Switching from torch._scaled_mm to vLLM's cutlass fp8 kernels when supported as we are seeing 5-15% improvement in e2e performance on neuralmagic/Meta-Llama-3-8B-Instruct-FP8

see https://docs.google.com/spreadsheets/d/1GiAnmzyGHgZ6zL_LDSTm35Bdrt4A8AaFEurDlISYYA4/ for some quick e2e benchmarks and #5144 for comparisons across different GEMM sizes.

The 2nd PR for #4532.

This PR supports loading FP8 kv-cache scaling factors from a FP8 checkpoint (with .kv_scale parameter).

This PR improves the FP8 performance of linear layers, which had been lacking before (#4118 (comment) and #4118 (comment)).

We noticed that CUBLASLt can find a better algorithm if the first dimension of the matrix is greater than 16. So this PR enlarges matrices appropriately during quantization. This improves FP8 performance and removes the performance regression vs. FP16, in many cases exceeding FP16 performance.

Here are benchmarks on llama3 70b (ITL numbers for 1000 input and 50 output tokens at fixed qps and at TP 4), all FP8 measurements are for dynamic quantization:

qps = 1: 24 ms (FP8, this PR), 32 ms (FP8, previous main), 26 ms (FP16)
qps = 2: 26 ms (FP8, this PR), 34ms (FP8, previous main), 28 ms (FP16) 
qps = 4: 33 ms (FP8, this PR), 44 ms (FP8, previous main), 36 ms (FP16)
qps = 6: 46 ms (FP8, this PR), 56 ms (FP8, previous main), 54 ms (FP16)
qps = 8: 85 ms (FP8, this PR), 85 ms (FP8, previous main), 138 ms (FP16)

Co-authored-by: Philipp Moritz <pcmoritz@gmail.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: mgoin <michael@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

Fixes fp8 iterface which broke in AQLM merge.

Provide an initial support to FP8 computation. This PR is inspired by HuggingFace TGI: huggingface/text-generation-inference#1726

This feature can be enabled with --quantization fp8 or -q fp8 when launching an engine.

Algorithm:
We still load a model checkpoint in FP16/BF16. After the weights are loaded, Fp8LinearMethod calculates the per-tensor scaling factor of weights and quantizes the weights accordingly. The scaling factor will then be stored for future use. Meanwhile, the per-tensor scaling factor for activations is calculated in every forward pass.

Initial Results:
Currently tested Mistral-7B on 1xH100. With prompt length ~5 and decoding length 128:

BF16: 1.47s
FP8: 1.66s
I'll try to use larger models and try to find more performance bottleneck. Meanwhile, you're welcome to try this code.