Issue Station
Sign In
Tools hereToolbox

vllm - 💡(How to fix) Fix [Feature]: Quantization support (AWQ / GPTQ / FP8) for mistralai/Voxtral-Mini-4B-Realtime-2602 [1 comments, 2 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#38235•Fetched 2026-04-08 01:37:06
View on GitHub
Comments
1
Participants
2
Timeline
2
Reactions
0
Author
sh1mansh1man
Participants
EricccYangEricccYangsh1mansh1man
Timeline (top)
commented ×1labeled ×1

Root Cause

I am currently deploying the new mistralai/Voxtral-Mini-4B-Realtime-2602 model using the vLLM V1 Engine. Because it is a realtime multimodal audio model, the memory footprint is extremely high.

RAW_BUFFERClick to expand / collapse

🚀 The feature, motivation and pitch

I am currently deploying the new mistralai/Voxtral-Mini-4B-Realtime-2602 model using the vLLM V1 Engine. Because it is a realtime multimodal audio model, the memory footprint is extremely high.

On a standard mid-tier GPU like the RTX 5060 Ti (16GB VRAM), the bf16 weights consume around 8.4GB. This leaves very little room for the encoder_cache and standard KV cache. As a result, the max-model-len has to be severely restricted (around ~4300 tokens), which is barely enough for a short audio session, and it completely prevents running multiple concurrent streams due to encoder_cache_usage hitting 100%.

It would be incredibly helpful to have quantization support for the VoxtralRealtimeGeneration architecture. Specifically:

  1. Weight Quantization: Support for loading AWQ, GPTQ, or bitsandbytes (int8/int4) formats for this specific model type.
  2. KV Cache Quantization: Support for FP8 KV cache for Voxtral to further reduce the memory required for long audio streams.

Reducing the weight footprint by half (e.g., to ~4GB using 8-bit) would drastically increase the viable audio context window and allow continuous batching of multiple sessions on 16GB and 24GB consumer GPUs.

Currently, the only alternative is to limit max-model-len to minimal values or use expensive multi-GPU setups / A100s, which makes testing and local deployment of Realtime audio models inaccessible for many developers.

Alternatives

No response

Additional context

No response

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 address the issue of high memory footprint in the mistralai/Voxtral-Mini-4B-Realtime-2602 model, we will implement weight quantization and KV cache quantization.

Weight Quantization

We will use the bitsandbytes library to support int8 weight quantization. Here's an example code snippet:

import bitsandbytes as bnb

# Load the model weights
weights = ...

# Quantize the weights to int8
quantized_weights = bnb.quantize(weights, symmetric=True)

# Save the quantized weights
torch.save(quantized_weights, 'quantized_weights.pth')

KV Cache Quantization

We will use FP8 quantization for the KV cache. Here's an example code snippet:

import torch

# Define the FP8 quantization function
def fp8_quantize(x):
    return torch.quantize_per_tensor(x, scale=1.0, zero_point=0, dtype=torch.qint8)

# Quantize the KV cache
kv_cache = ...
quantized_kv_cache = fp8_quantize(kv_cache)

Verification

To verify that the fix worked, we can check the memory usage of the model before and after applying the quantization. We can use the nvidia-smi command to monitor the GPU memory usage.

Extra Tips

  • Make sure to update the model loading code to support the quantized weight formats.
  • Test the model with different input sizes to ensure that the quantization does not affect the model's performance.
  • Consider using other quantization libraries, such as torch.quantization, for more advanced quantization techniques.

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#GPU setup#API versioning#request timeout

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 [Feature]: Quantization support (AWQ / GPTQ / FP8) for mistralai/Voxtral-Mini-4B-Realtime-2602 [1 comments, 2 participants]