Skip to content

[plugin][script][recipe] update env vars for kimi and minimax #1387

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

Open
gbyu-amd wants to merge 2 commits into
base: main
Choose a base branch
from
+16 −10
Open

[plugin][script][recipe] update env vars for kimi and minimax #1387

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b789a17
ci(benchmark): enable flydsl MoE sort + AR 1-stage cap for Kimi-K2.5 …
ganyi1996ppo Jun 28, 2026
1392c13
docs(recipes): mirror Kimi-K2.5 / MiniMax-M2.5 env vars in atom-vllm …
ganyi1996ppo Jun 28, 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
12 changes: 6 additions & 6 deletions .github/benchmark/oot_benchmark_models.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -164,7 +164,7 @@
"prefix": "kimi-k25-mxfp4-tp4-met",
"bench_args": "",
"extra_args": "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512"
},
{
"tp_size": 4,
Expand All @@ -173,7 +173,7 @@
"prefix": "kimi-k2-5-mxfp4-aw-tp4",
"bench_args": "",
"extra_args": "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512"
},
{
"tp_size": 8,
Expand All @@ -182,7 +182,7 @@
"prefix": "kimi-k2-5-mxfp4-aw-tp8",
"bench_args": "",
"extra_args": "--trust-remote-code --tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512"
}
]
},
Expand Down Expand Up @@ -433,7 +433,7 @@
"prefix": "minimax-m2-5-aw-tp2",
"bench_args": "",
"extra_args": "--trust-remote-code --tensor-parallel-size 2 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1"
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048"
},
{
"tp_size": 4,
Expand All @@ -442,7 +442,7 @@
"prefix": "minimax-m2-5-aw-tp4",
"bench_args": "",
"extra_args": "--trust-remote-code --tensor-parallel-size 4 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1"
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048"
},
{
"tp_size": 8,
Expand All @@ -451,7 +451,7 @@
"prefix": "minimax-m2-5-aw-tp8",
"bench_args": "",
"extra_args": "--trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1"
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048"
}
]
}
Expand Down
8 changes: 4 additions & 4 deletions .github/benchmark/oot_models_accuracy.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,7 +165,7 @@
"model_name": "Kimi-K2.5-MXFP4 TP4",
"model_path": "amd/Kimi-K2.5-MXFP4-AttnFP8",
"extra_args": "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512",
"runner": "atom-plugin-acc-validation-runner",
"test_level": "nightly",
"priority": "P0",
Expand All @@ -175,7 +175,7 @@
"model_name": "Kimi-K2.5-MXFP4 TP8",
"model_path": "amd/Kimi-K2.5-MXFP4-AttnFP8",
"extra_args": "--trust-remote-code --tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512",
"runner": "atom-plugin-acc-validation-runner",
"test_level": "nightly",
"priority": "P1",
Expand Down Expand Up @@ -310,7 +310,7 @@
"model_name": "MiniMax-M2.5 TP2",
"model_path": "MiniMaxAI/MiniMax-M2.5",
"extra_args": "--trust-remote-code --tensor-parallel-size 2 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048",
"runner": "atom-plugin-acc-validation-runner",
"test_level": "nightly",
"priority": "P0",
Expand All @@ -320,7 +320,7 @@
"model_name": "MiniMax-M2.5 TP4",
"model_path": "MiniMaxAI/MiniMax-M2.5",
"extra_args": "--trust-remote-code --tensor-parallel-size 4 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1",
"env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048",
"runner": "atom-plugin-acc-validation-runner",
"test_level": "nightly",
"priority": "P1",
Expand Down
4 changes: 4 additions & 0 deletions recipes/atom_vllm/Kimi-K2.5.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,10 @@ We adopt [amd/Kimi-K2.5-MXFP4-AttnFP8](https://huggingface.co/amd/Kimi-K2.5-MXFP
```bash
# use quick allreduce to reduce TTFT
export AITER_QUICK_REDUCE_QUANTIZATION=INT4
# use flydsl MoE sorting kernel
export AITER_USE_FLYDSL_MOE_SORTING=1
# cap the single-stage allreduce size (KB) to favor the faster path
export AITER_AR_1STAGE_MAX_KB=512

vllm serve amd/Kimi-K2.5-MXFP4-AttnFP8 \
--host localhost \
Expand Down
2 changes: 2 additions & 0 deletions recipes/atom_vllm/MiniMax-M2.5.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,8 @@ The following matches the vLLM-ATOM benchmark entries in `.github/benchmark/oot_
export AITER_QUICK_REDUCE_QUANTIZATION=INT4
export ATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1
export ATOM_USE_GLUON_PA_DECODE=1
# cap the single-stage allreduce size (KB) to favor the faster path
export AITER_AR_1STAGE_MAX_KB=2048

vllm serve MiniMaxAI/MiniMax-M2.5 \
--host localhost \
Expand Down
Loading