minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[XPU][NIGHTLY] Add 8 XPU nightly tests, enable 1-gpu suite (#30246)

Browse Source 
Co-authored-by: arathi-hlab <arathi-hlab@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
ashwini rathi
2026-07-20 16:15:50 +08:00
committed by GitHub
co-authored by arathi-hlab Claude Opus 4.8
parent 50c118704a
commit fafa302e41
13 changed files with 828 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/test/xpu/simple_eval_gsm8k_xpu_mixin.py
+7 -16
View File
@@ -1,9 +1,8 @@
"""simple-evals GSM8K accuracy mixin for Intel XPU nightly tests.
Mirrors the AMD/NVIDIA nightly flow (``test_gsm8k_eval_amd.py`` /
``test_text_models_gsm8k_eval.py``): launch an SGLang server with XPU
flags, then call ``sglang.test.run_eval`` with ``eval_name="gsm8k"`` so
the same ``simple_eval_gsm8k.GSM8KEval`` evaluator scores every backend.
Launches an SGLang server with XPU flags, then calls ``sglang.test.run_eval``
with ``eval_name="gsm8k"`` so the ``simple_eval_gsm8k.GSM8KEval`` evaluator
scores the run.
Subclasses set ``model``, ``tp_size``, ``accuracy``, and may override
``other_args`` / ``env`` / ``num_examples`` / ``num_threads``.
@@ -46,20 +45,12 @@ class SimpleEvalGSM8KXPUMixin(ABC):
env: dict | None = None
server_cmd: str = ""
# 200 questions matches the limit used by the XPU 70B lm-eval YAML and
# fits inside run_suite's per-file timeout when num_threads=1 keeps
# throughput low. Subclasses on cheaper-per-token hardware (TP=1, no
# Level Zero wedge) can raise this or set None for the full 1319-question
# GSM8K test set, matching the AMD/NVIDIA nightly defaults.
# Subset that fits run_suite's per-file timeout; set None for the full set.
num_examples: int | None = 200
# Single-stream eval: intel_xpu attention at TP>=2 wedges the Level Zero
# driver in ur_command_list_manager::appendUSMMemcpy on concurrent prefill.
# Subclasses on hardware that handles parallel prefill cleanly may bump.
# Single-stream: intel_xpu attention at TP>=2 wedges the Level Zero driver
# on concurrent prefill. Subclasses may bump on hardware that handles it.
num_threads: int = 1
# Short generations reduce the rate of prefill->decode->prefill handoffs,
# which is what trips the same Level Zero wedge on TP>=2 (observed at the
# default 2048; 512 matches the original few_shot_gsm8k limit and is still
# enough for GSM8K CoT answers).
# Short generations reduce prefill->decode handoffs that trip the same wedge.
max_tokens: int = 512
@classmethod