[CI] Move compatible 1xL4 jobs to H200 35GB MIG (#43024)

Signed-off-by: Simon Mo <simon@inferact.ai>
Co-authored-by: Simon Mo <simon@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Kevin H. Luu
2026-07-19 19:21:25 -07:00
committed by GitHub
co-authored by Simon Mo OpenAI Codex OpenAI Codex
parent ace9fda495
commit 1dcbbd9cac
17 changed files with 129 additions and 15 deletions
@@ -188,6 +188,16 @@ def test_models(
prompt_embeds.append(embed.squeeze(0))
vllm_kwargs = {}
if (
model == "bigscience/bloom-560m"
and current_platform.is_device_capability_family(90)
):
# On SM90, the metadata builder otherwise selects FA3 AOT scheduling
# before Bloom's ALiBi layers fall back to FA2. Pinning FA2 keeps the
# builder and layer consistent and preserves the L4 test path.
vllm_kwargs["attention_config"] = {"flash_attn_version": 2}
with vllm_runner(
model,
tokenizer_name=model_info.tokenizer or model,
@@ -200,6 +210,7 @@ def test_models(
max_num_seqs=1 if current_platform.is_rocm() else 2,
enable_prompt_embeds=use_prompt_embeds,
compilation_config={"cudagraph_capture_sizes": [1, 2]},
**vllm_kwargs,
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs
@@ -384,8 +384,13 @@ def test_fp32_cache_state(
example_prompts, max_tokens, num_logprobs
)
# Leave enough headroom for repeated engine initialization on a
# 32.5 GiB MIG.
with vllm_runner(
model, max_num_seqs=MAX_NUM_SEQS, **{cache_dtype_param: "float32"}
model,
max_num_seqs=MAX_NUM_SEQS,
gpu_memory_utilization=0.9,
**{cache_dtype_param: "float32"},
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs
+2
View File
@@ -67,6 +67,8 @@ def test_models(
if kv_cache_dtype == "fp8_e5m2" and current_platform.is_rocm():
pytest.skip(f"{kv_cache_dtype} is currently not supported on ROCm/HIP.")
if kv_cache_dtype == "fp8_e5m2" and current_platform.is_cuda():
pytest.skip(f"{kv_cache_dtype} is not supported by FLASH_ATTN on CUDA.")
if not (
current_platform.is_xpu()