Skip to content

[feat](ci): add glm5.1 and dsv4-flash on sglang #1413

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
valarLip merged 2 commits into from
Jun 30, 2026
+55 −2
Merged

[feat](ci): add glm5.1 and dsv4-flash on sglang #1413

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1dd7831
[feat](ci): add glm5.1 and dsv4-flash on sglang
zgplvyou Jun 30, 2026
f223e89
[fix](ci): add label for ds
zgplvyou Jun 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 .github/benchmark/oot_models_accuracy.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -201,7 +201,7 @@
"model_path": "deepseek-ai/DeepSeek-R1-0528",
"extraArgs": "--data-parallel-size 8 --enable-expert-parallel",
"env_vars": "MORI_SHMEM_MODE=ISOLATION",
"runner": "linux-atom-mi35x-8",
"runner": "atom-plugin-acc-validation-runner",
Comment thread
junyyang-amd marked this conversation as resolved.
"test_level": "nightly",
"accuracy_threshold": 0.93,
"accuracy_baseline": 0.93,
Expand Down
28 changes: 28 additions & 0 deletions .github/benchmark/sglang_models_accuracy.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -326,5 +326,33 @@
"accuracy_baseline": null,
"accuracy_baseline_model": "Qwen/Qwen3-32B-FP8",
"_baseline_note": "Threshold placeholder until MI308 gsm8k baseline is measured in CI."
},
{
"model_name": "MI308 GLM-5.1-FP8 TP8",
"model_path": "zai-org/GLM-5.1-FP8",
"extraArgs": "--trust-remote-code --tensor-parallel-size 8 --attention-backend aiter --kv-cache-dtype fp8_e4m3 --mem-fraction-static 0.8 --page-size 1 --disable-radix-cache",
"env_vars": "SGLANG_AITER_FP8_PREFILL_ATTN=0\nSGLANG_USE_AITER=1\nATOM_ENABLE_DS_QKNORM_QUANT_FUSION=1\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nSGLANG_ENABLE_TORCH_COMPILE=1\nTORCHINDUCTOR_COMPILE_THREADS=128",
"runner": "atom-mi308-8gpu-plugins-benchmark",
"test_level": "nightly",
"lm_eval_num_fewshot": 20,
"lm_eval_num_concurrent": 64,
"accuracy_threshold": 0.93,
"accuracy_baseline": null,
"accuracy_baseline_model": "zai-org/GLM-5.1-FP8",
"_baseline_note": "Threshold aligned with the SGLANG accuracy validation workflow target for GLM-5.1-FP8 TP8 on MI308."
},
{
"model_name": "MI308 DeepSeek-V4-Flash",
"model_path": "deepseek-ai/DeepSeek-V4-Flash",
"extraArgs": "--trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8_e4m3 --mem-fraction-static 0.9 --swa-full-tokens-ratio 0.1 --max-running-requests 256 --page-size 256 --disable-radix-cache --disable-shared-experts-fusion --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4",
"env_vars": "SGLANG_DEFAULT_SERVER_ARGS=\nAITER_BF16_FP8_MOE_BOUND=0\nATOM_MOE_GU_ITLV=1\nSGLANG_DEFAULT_THINKING=1\nSGLANG_DSV4_REASONING_EFFORT=max\nSGLANG_USE_AITER=1\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nTORCHINDUCTOR_COMPILE_THREADS=128",
"runner": "atom-mi308-8gpu-plugins-benchmark",
"test_level": "nightly",
"lm_eval_num_fewshot": 5,
"lm_eval_num_concurrent": 8,
"accuracy_threshold": 0.94,
"accuracy_baseline": null,
"accuracy_baseline_model": "deepseek-ai/DeepSeek-V4-Flash",
"_baseline_note": "MI308 SGLang DeepSeek-V4-Flash coverage follows the SGLang DeepSeek-V4 5-shot GSM8K setup; refresh baseline after nightly measurements land."
}
]
4 changes: 4 additions & 0 deletions .github/workflows/atom-sglang-accuracy-validation-gpu-shard.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -350,6 +350,8 @@ jobs:
MAX_WAIT_RETRIES: "120"
STREAM_SGLANG_LOGS: "1"
LM_EVAL_TASK: "gsm8k"
LM_EVAL_NUM_FEWSHOT: ${{ matrix.lm_eval_num_fewshot }}
LM_EVAL_NUM_CONCURRENT: ${{ matrix.lm_eval_num_concurrent }}
run: |
$CONTAINER_ENGINE exec \
-e SGLANG_MODEL_NAME="${SGLANG_MODEL_NAME}" \
Expand All @@ -363,6 +365,8 @@ jobs:
-e MAX_WAIT_RETRIES="${MAX_WAIT_RETRIES}" \
-e STREAM_SGLANG_LOGS="${STREAM_SGLANG_LOGS}" \
-e LM_EVAL_TASK="${LM_EVAL_TASK}" \
-e LM_EVAL_NUM_FEWSHOT="${LM_EVAL_NUM_FEWSHOT}" \
-e LM_EVAL_NUM_CONCURRENT="${LM_EVAL_NUM_CONCURRENT}" \
"$CONTAINER_NAME" bash -lc "
set -euo pipefail
bash .github/scripts/atom_sglang_test.sh accuracy
Expand Down
22 changes: 21 additions & 1 deletion .github/workflows/atom-sglang-accuracy-validation.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,7 +120,7 @@ on:
type: boolean
default: false
run_mi308_all:
description: "Run all MI308 gsm8k accuracy cases (Qwen on atom-mi308-8gpu-plugins-benchmark)"
description: "Run all MI308 gsm8k accuracy cases on atom-mi308-8gpu-plugins-benchmark"
required: false
type: boolean
default: false
Expand Down Expand Up @@ -448,6 +448,26 @@ jobs:
"env_vars": "SGLANG_DEFAULT_SERVER_ARGS=\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0",
"runner": "atom-mi308-8gpu-plugins-benchmark",
},
{
"model_name": "MI308 GLM-5.1-FP8 TP8",
"model_path": "zai-org/GLM-5.1-FP8",
"extra_args": "--trust-remote-code --tensor-parallel-size 8 --attention-backend aiter --kv-cache-dtype fp8_e4m3 --mem-fraction-static 0.8 --page-size 1 --disable-radix-cache",
"lm_eval_num_fewshot": 20,
"lm_eval_num_concurrent": 64,
"accuracy_test_threshold": 0.93,
"env_vars": "SGLANG_AITER_FP8_PREFILL_ATTN=0\nSGLANG_USE_AITER=1\nATOM_ENABLE_DS_QKNORM_QUANT_FUSION=1\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nSGLANG_ENABLE_TORCH_COMPILE=1\nTORCHINDUCTOR_COMPILE_THREADS=128",
"runner": "atom-mi308-8gpu-plugins-benchmark",
},
{
"model_name": "MI308 DeepSeek-V4-Flash",
"model_path": "deepseek-ai/DeepSeek-V4-Flash",
"extra_args": "--trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8_e4m3 --mem-fraction-static 0.9 --swa-full-tokens-ratio 0.1 --max-running-requests 256 --page-size 256 --disable-radix-cache --disable-shared-experts-fusion --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4",
"lm_eval_num_fewshot": 5,
"lm_eval_num_concurrent": 8,
"accuracy_test_threshold": 0.94,
"env_vars": "SGLANG_DEFAULT_SERVER_ARGS=\nAITER_BF16_FP8_MOE_BOUND=0\nATOM_MOE_GU_ITLV=1\nSGLANG_DEFAULT_THINKING=1\nSGLANG_DSV4_REASONING_EFFORT=max\nSGLANG_USE_AITER=1\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nTORCHINDUCTOR_COMPILE_THREADS=128",
"runner": "atom-mi308-8gpu-plugins-benchmark",
},
]
models.extend(mi308_models)

Expand Down
Loading