[CI] Move tests that do not need GB300 off the per-PR GB300 suite

[mmangkad]

Summary

Run tests that do not depend on GB300 hardware on other runners, and replace one end-to-end class with CPU unit tests.

Test	From	To
kernels/ops/attention/test_flash_attention_4_sm103.py	4-gpu-gb300 base-c	8-gpu-b300 base-c (B300 is also SM103)
kernels/ops/attention/test_kda_prefill.py	4-gpu-gb300 base-c	8-gpu-b300 base-c; the 4-gpu-b200 registration is unchanged
ep/test_flashinfer_a2a.py	4-gpu-gb300 base-c	4-gpu-b200 nightly; GB300-specific NCCL port pin removed; GLM-5.2 static-dispatch class replaced by unit tests (below)
e2e/models_large/test_deepseek_v3_cutedsl_4gpu.py	4-gpu-gb300 base-c	4-gpu-b200 nightly; GB300-specific NCCL port pin removed
e2e/models/test_deepseek_v41_flash_gb300.py	est_time=1800	est_time=1100, the measured runtime

disaggregation_aarch64, cuda_vmm_transport and numa_utils stay on 4-gpu-gb300 base-c.

Why nightly is enough for the two moved end-to-end tests: the components they exercise are already checked per-PR:

Component	Per-PR test
FlashInfer A2A dispatcher	unit/layers/moe/test_flashinfer_dispatcher.py, unit/layers/moe/test_flashinfer_a2a_wide_ep.py
NVFP4 MoE runners (CuteDSL, TRT-LLM)	unit/layers/quantization/test_nvfp4_moe_backends.py
DeepEP dispatch	e2e/models/test_deepseek_v4_flash_fp4_b200.py, ep/test_deepep_small.py

These two files add end-to-end accuracy across models, which does not need per-PR cadence.

FlashInfer A2A top-k unit tests

#29929 fixed two things in FlashinferDispatcher, which feeds FlashInfer's MoeAlltoAll:

• it casts expert IDs to int32, because static EP dispatch produces int64 and MoeAlltoAll rejects it;
• it materializes bypassed top-k output, from runners that fuse routing, before dispatch.

The GLM-5.2 class in test_flashinfer_a2a.py covered only the first, by launching a full GLM-5.2-NVFP4 server and asserting a non-empty /generate response. It used the CuteDSL runner, which produces standard top-k, so the second fix had no test.

Two cases in TestFlashinferDispatcher (test_flashinfer_dispatcher.py) now cover both on CPU, per-PR, with a recording MoeAlltoAll:

Case	Asserts	With that part of the fix removed
test_int64_expert_ids_reach_moe_a2a_as_int32	int64 IDs arrive as int32 (argument and expert-ID payload), values unchanged	AssertionError: torch.int64 != torch.int32
test_bypassed_topk_is_routed_before_moe_a2a	a bypassed top-k output arrives as routed int32 expert IDs plus weights	AttributeError: 'BypassedTopKOutput' object has no attribute 'topk_ids'

The file's existing case now lives in the same CustomTestCase class, with its assertions unchanged.

FlashInfer A2A with the CuteDSL runner stays covered end to end by moe/test_flashinfer_a2a_cutedsl_v2.py (extra-b) and e2e/models_large/test_nvidia_nemotron_3_super_nvfp4.py (nightly). The GLM class used the standard top-k format, so no bypassed-format coverage is lost. test_flashinfer_a2a.py drops from about 544s to about 360s.

Shard counts

From compute_partitions with the live timing model:

Suite	main	this branch
base-c-test-4-gpu-gb300	3	1
base-c-test-8-gpu-b300	1	1
nightly-test-4-gpu-b200	2	2

Test plan

• pre-commit, including the registered-test CI registry check
• test_flashinfer_dispatcher.py on CPU: 3 passed; each new case fails with its part of the fix removed
• 4-gpu-b200: test_flashinfer_a2a.py, test_deepseek_v3_cutedsl_4gpu.py (run)
• 8-gpu-b300: test_flash_attention_4_sm103.py, test_kda_prefill.py (run)

---

CI States

Latest PR Test (Base): ❌ Run #38213861067
Latest PR Test (Extra): ❌ Run #38213860851
Latest PR Test (AMD ROCm 10): ❌ Run #38213861027

[mmangkad]

/rerun-test test/registered/e2e/models/test_deepseek_v41_flash_gb300.py test/registered/e2e/models_large/test_deepseek_v3_cutedsl_4gpu.py test/registered/ep/test_flashinfer_a2a.py test/registered/kernels/ops/attention/test_flash_attention_4_sm103.py test/registered/kernels/ops/attention/test_kda_prefill.py test/registered/unit/layers/moe/test_flashinfer_dispatcher.py

[github-actions]

🚀 4-gpu-gb300 (1 test): ✅ View workflow run

cd test/ && python3 registered/e2e/models/test_deepseek_v41_flash_gb300.py

🚀 4-gpu-b200 (3 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/models_large/test_deepseek_v3_cutedsl_4gpu.py
cd test/ && python3 registered/ep/test_flashinfer_a2a.py
cd test/ && python3 registered/kernels/ops/attention/test_kda_prefill.py

🚀 8-gpu-b300 (2 tests): ✅ View workflow run

cd test/ && python3 registered/kernels/ops/attention/test_flash_attention_4_sm103.py
cd test/ && python3 registered/kernels/ops/attention/test_kda_prefill.py

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/layers/moe/test_flashinfer_dispatcher.py