GitLab

Currently we need to call rotary embedding kernel for each LoRA, which makes it hard to serve multiple long context length LoRA. Add batched rotary embedding kernel and pipe it through.

It replaces the rotary embedding layer to the one that is aware of multiple cos-sin-cache per scaling factors.

Follow up of https://github.com/vllm-project/vllm/pull/3095/files

Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>

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

[Core][2/N] Model runner refactoring part 2. Combine prepare prefill / decode to a single API (#4681)

This PR combines prepare_prompt and prepare_decode into a single API. This PR also coelsce the attn metadata for prefill/decode to a single class and allow to slice them when running attn backend.

It also refactors subquery_start_loc which was not refactored in the previous PR

[Frontend] [Core] perf: Automatically detect vLLM-tensorized model, update `tensorizer` to version 2.9.0 (#4208)

Storing exception frame is extremely prone to circular refernece because it contains the reference to objects.

When tensorizer is not installed, it leaks llm instance because error frame has references to various modules which cause circular reference problem.

I also found spec decoding has a circular reference issue, and I solved it using weakref.proxy.

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: Dash Desai <1723932+iamontheinet@users.noreply.github.com>
Co-authored-by: Aurick Qiao <qiao@aurick.net>
Co-authored-by: Aurick Qiao <aurick.qiao@snowflake.com>
Co-authored-by: Aurick Qiao <aurickq@users.noreply.github.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

Co-authored-by: Cade Daniel <edacih@gmail.com>

[Kernel] Support MoE Fp8 Checkpoints for Mixtral (Static Weights with Dynamic/Static Activations) (#4527)

Follow on to #4332 to enable FP8 checkpoint loading for Mixtral and supersedes #4436.

This PR enables the following checkpoint loading features for Mixtral:

Supports loading fp8 checkpoints for Mixtral, such as this "nm-testing/Mixtral-8x7B-Instruct-v0.1-FP8" test model
Supports static or dynamic activation quantization with static weight quantization (all per tensor)
Supports different scales for each expert weight
Supports Fp8 in QKV layer
Notes:

The Expert Gate/Router always runs at half / full precision for now.
If there are different weight scales between QKV layer (for separate QKV weights), they are re-quantized using layer.weight_scale.max() so we can have a single gemm for performance.

Remove the device="cuda" declarations in mixtral as promised in #4343

This PR updates the tuning script for the fused_moe kernel to support FP8 and also adds configurations for TP4. Note that for the configuration I removed num_warps and num_stages for small batch sizes since that improved performance and brought the benchmarks on par with the numbers before in that regime to make sure this is a strict improvement over the status quo.

All the numbers below are for mistralai/Mixtral-8x7B-Instruct-v0.1, 1000 input and 50 output tokens.

Before this PR (with static activation scaling):

qps = 1: 9.8 ms ITL, 0.49s e2e latency
qps = 2: 9.7 ms ITL, 0.49s e2e latency 
qps = 4: 10.1 ms ITL, 0.52s e2e latency
qps = 6: 11.9 ms ITL, 0.59s e2e latency
qps = 8: 14.0 ms ITL, 0.70s e2e latency
qps = 10: 15.7 ms ITL, 0.79s e2e latency

After this PR (with static activation scaling):

qps = 1: 9.8 ms ITL, 0.49s e2e latency
qps = 2: 9.7 ms ITL, 0.49s e2e latency
qps = 4: 10.2 ms ITL, 0.53s e2e latency
qps = 6: 11.9 ms ITL, 0.59s e2e latency
qps = 8: 11.9 ms ITL, 0.59s e2e latency
qps = 10: 12.1 ms ITL, 0.61s e2e latency

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: alexm <alexm@neuralmagic.com>
Co-authored-by: mgoin <michael@neuralmagic.com>

Signed-off-by: Prashant Gupta <prashantgupta@us.ibm.com>
Co-authored-by: Travis Johnson <tjohnson31415@gmail.com>

Co-authored-by: DefTruth <31974251+deftruth@users.noreply.github.com>