Skip to content

[Bug]: RuntimeError: Shared memory exceeds 96KB: 146944 bytes in MiniMax-M2.7 #40

@bayley

Description

@bayley

Your current environment

8x V100 32GB

🐛 Describe the bug

#!/bin/bash

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

vllm serve /home/user/models/QuantTrio--MiniMax-M2.7-AWQ --served-model-name minimax-m2.7 \
  --dtype float16 \
  --gpu-memory-utilization 0.9 \
  --tensor-parallel-size 8 \
  --max-num-seqs 4 \
  --max-num-batched-tokens 4096 \
  --attention-backend FLASH_ATTN_V100 \
  --reasoning-parser minimax_m2 \
  --enable-auto-tool-choice \
  --tool-call-parser minimax_m2 \
  --trust-remote-code

There have clearly been some major changes in the Flash Attention codebase, because now decode fails with RuntimeError: Shared memory exceeds 96KB: 146944 bytes in MiniMax-M2.7. If I vaguely recall correctly, this happens when FA tries to allocate shared memory that exceeds the available size per SM, which occurs when head dimensions exceed some size.

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.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions