Issue Station
Sign In
Tools hereToolbox

vllm - 💡(How to fix) Fix [Performance]: W4A16+eagle3 not better than fp8+eagle3 with Qwen2.5-14B [1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
vllm-project/vllm#36629•Fetched 2026-04-08 00:35:49
View on GitHub
Comments
0
Participants
1
Timeline
1
Reactions
0
Author
ljy6j13ljy6j13
Participants
ljy6j13ljy6j13
Timeline (top)
labeled ×1

Code Example

4090d * 1
vllm 0.16.0
fp8 + eagle3 extra information ：CutlassFP8ScaledMMLinearKernel
w4a16 + eagle3 extra information: MarlinLinearKernel
RAW_BUFFERClick to expand / collapse

Proposal to improve performance

No response

Report of performance regression

Under low concurrency (1-8), the performance of w4a16 + eagle3 is higher than that of fp8 + eagle3. However, under high concurrency (16-32), it seems that there is a conflict between w4a16 and eagle3, and the performance is even worse than that of fp8 + eagle3.

fp8+eagle3 batch16 ============ Serving Benchmark Result ============ Successful requests: 500 Failed requests: 0 Maximum request concurrency: 16 Request rate configured (RPS): 50.00 Benchmark duration (s): 40.97 Total input tokens: 84383 Total generated tokens: 46345 Request throughput (req/s): 12.20 Output token throughput (tok/s): 1131.19 Peak output token throughput (tok/s): 464.00 Peak concurrent requests: 35.00 Total Token throughput (tok/s): 3190.81 ---------------Time to First Token---------------- Mean TTFT (ms): 121.16 Median TTFT (ms): 115.09 P95 TTFT (ms): 172.56 P99 TTFT (ms): 185.92 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 13.16 Median TPOT (ms): 12.78 P95 TPOT (ms): 16.98 P99 TPOT (ms): 19.36 ---------------Inter-token Latency---------------- Mean ITL (ms): 35.81 Median ITL (ms): 30.46 P95 ITL (ms): 56.24 P99 ITL (ms): 68.16

w4a16+eagle3 batch16 ============ Serving Benchmark Result ============ Successful requests: 500 Failed requests: 0 Maximum request concurrency: 16 Request rate configured (RPS): 50.00 Benchmark duration (s): 42.96 Total input tokens: 84383 Total generated tokens: 46362 Request throughput (req/s): 11.64 Output token throughput (tok/s): 1079.15 Peak output token throughput (tok/s): 480.00 Peak concurrent requests: 33.00 Total token throughput (tok/s): 3043.31 ---------------Time to First Token---------------- Mean TTFT (ms): 135.96 Median TTFT (ms): 129.76 P95 TTFT (ms): 197.06 P99 TTFT (ms): 232.51 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 13.73 Median TPOT (ms): 13.38 P95 TPOT (ms): 18.56 P99 TPOT (ms): 21.28 ---------------Inter-token Latency---------------- Mean ITL (ms): 37.33 Median ITL (ms): 30.69 P95 ITL (ms): 73.98 P99 ITL (ms): 95.74 ---------------Speculative Decoding--------------- Acceptance rate (%): 62.22 Acceptance length: 2.87 Drafts: 16254 Draft tokens: 48762 Accepted tokens: 30341 Per-position acceptance (%): Position 0: 76.97 Position 1: 60.75 Position 2: 48.95

What could be the possible reasons, and is there any plan for kernel optimization for w4a16?

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

4090d * 1
vllm 0.16.0
fp8 + eagle3 extra information ：CutlassFP8ScaledMMLinearKernel
w4a16 + eagle3 extra information: MarlinLinearKernel

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

extent analysis

Fix Plan

To improve the performance of w4a16 + eagle3 under high concurrency, we need to optimize the kernel.

Here are the steps:

  • Update the MarlinLinearKernel to use a more efficient algorithm, such as the CutlassFP8ScaledMMLinearKernel used in fp8 + eagle3.
  • Increase the batch size to reduce the overhead of kernel launches.
  • Use a more efficient data type, such as float16, to reduce memory bandwidth usage.

Example code:

import torch

# Update the kernel
class OptimizedMarlinLinearKernel(torch.nn.Module):
    def __init__(self):
        super(OptimizedMarlinLinearKernel, self).__init__()
        self.linear = torch.nn.Linear(128, 128, bias=False)

    def forward(self, x):
        return self.linear(x)

# Increase the batch size
batch_size = 32

# Use a more efficient data type
dtype = torch.float16

Verification

To verify that the fix worked, run the benchmark again with the optimized kernel and increased batch size. Check the request throughput, output token throughput, and inter-token latency to ensure they have improved.

Extra Tips

  • Monitor the GPU utilization and memory usage to ensure the optimized kernel is not causing any bottlenecks.
  • Consider using a more efficient optimizer, such as AdamW, to further improve performance.
  • Test the optimized kernel with different batch sizes and data types to find the optimal configuration.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#ssr#installation#tensor shape#optimization#device allocation#model download#tokenizer error#prompt formatting

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

vllm - 💡(How to fix) Fix [Performance]: W4A16+eagle3 not better than fp8+eagle3 with Qwen2.5-14B [1 participants]