Skip to content

Add support for Mistral Large 3 inference with Flashinfer MoE #33174

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
vllm-bot merged 10 commits into from
Jan 31, 2026
Merged

Add support for Mistral Large 3 inference with Flashinfer MoE #33174

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
487151d
Add support for Mistral Large 3 inference with Flashinfer MoE
dbari Jan 26, 2026
389f2ea
Bump flashinfer version in docker
dbari Jan 29, 2026
186aecc
Update comment
dbari Jan 29, 2026
4e5484d
Extend assert
dbari Jan 29, 2026
026b4a6
Revert unneeded change
dbari Jan 29, 2026
463b08d
Add Triton configs for H200 TP8 with/without EP
dbari Jan 29, 2026
eb1dd0f
Extend benchmark_moe.py for Mistral Large 3
dbari Jan 29, 2026
56830ee
Merge branch 'main' into mistral_large_3_blackwell_blockwise
DarkLight1337 Jan 30, 2026
a135959
Fix test_flashinfer_per_tensor_moe_fp8_no_graph
dbari Jan 30, 2026
6fa035d
Merge remote-tracking branch 'origin/main' into mistral_large_3_black…
dbari Jan 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion requirements/cuda.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,4 +10,4 @@ torchaudio==2.9.1
# These must be updated alongside torch
torchvision==0.24.1 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
# FlashInfer should be updated together with the Dockerfile
flashinfer-python==0.6.1
flashinfer-python==0.6.2
Comment thread
dbari marked this conversation as resolved.
147 changes: 147 additions & 0 deletions ...executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_B200,dtype=fp8_w8a8.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,147 @@
{
"triton_version": "3.4.0",

@mgoin mgoin Jan 29, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I will note that these triton versions do seem out of date for modern torch+triton. It should be triton==3.5.1 for what we use on main

@dbari dbari Jan 29, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good catch, this was generated a while ago. I'll see if anything changes if I run the benchmark in the current environment.

@dbari dbari Jan 30, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it be ok to leave these as they are? It would take quite a bit of time to regenerate. Also, please keep in mind that the older ones are for the per-tensor FP8 quantization, which is only used in the Eagle draft model. The main model uses blockwise quantization and those configurations are newer (3.5.x).

"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
}
}
147 changes: 147 additions & 0 deletions vllm/model_executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_B200.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,147 @@
{
"triton_version": "3.4.0",
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 5
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 3
}
}
Loading