[Bugfix][ROCm] Fix stale model_config test stub breaking MoE multi-GPU combos; add mi300_2 CI job

[divakar-amd]

This PR does 2 things:

1. Fixes a multi-gpu error in mk
2. Adds mi300_2 mirror to run multi-gpu mk tests

Why

Running VLLM_ROCM_USE_AITER=1 pytest tests/kernels/moe/test_modular_kernel_combinations.py -k "aiter" on an MI300 (2-GPU) box, test_modular_kernel_combinations_multigpu[...] died with:

AttributeError: 'types.SimpleNamespace' object has no attribute 'sleep_mode_offload_cudagraph'

tests/kernels/moe/modular_kernel_tools/common.py fakes vllm_config.model_config as a bare SimpleNamespace(enforce_eager=True, is_moe=True). #59160 later added VllmConfig.use_cumem_cudagraph_pool, which reads model_config.sleep_mode_offload_cudagraph / .enable_sleep_mode / .sleep_mode_backend — fields the stub predates. That property is evaluated in CudaCommunicator.__init__ whenever a GroupCoordinator with world_size > 1 is built under set_current_vllm_config, which the multi-GPU DP test path does.

This fixes the stub by adding the three fields with the same defaults as real ModelConfig (False, False, "cumem").

Why this wasn't caught by existing CI

The only AMD CI job covering this file, ":amd: (MI300) MoE Kernels Shard %N", runs on the single-GPU mi300_1 agent pool. test_modular_kernel_combinations_multigpu always needs world_size=2, so it self-skips via its own current_platform.device_count() < world_size check before ever reaching the broken code path — the job shows green while never actually exercising the multi-GPU combinations.

This is the same bug class as #53840 / #53925 (ROCm CI jobs under-provisioned relative to a test's declared GPU requirement), but this specific instance wasn't in that audit's list — their scan matched the standard "Need at least N GPUs to run the test." skip message, while this test uses a differently-worded custom message ("Not enough GPUs available to run, got ... expected ..."), so it didn't show up in their log grep.

What's added

• tests/kernels/moe/modular_kernel_tools/common.py: add the three missing sleep_mode_* fields to the model_config stub.
• .buildkite/test-amd.yaml: new ":amd: (MI300) MoE Kernels Multi-GPU" job on mi300_2 (num_gpus: 2) running just the -k "multigpu" cases, so they actually execute instead of perpetually skipping. The existing mi300_1 job now deselects multigpu cases (-k "not multigpu") so the two jobs don't both claim the same test IDs — mirrors the "carved out, nothing double-runs" pattern from [ROCm][CI] Add multi-GPU test groups for nightly-log-confirmed skips (#53840) #53925.

Test commands run and results

Full suite, MI300 (2 GPUs), before and after the fix:

VLLM_ROCM_USE_AITER=1 pytest -s -v tests/kernels/moe/test_modular_kernel_combinations.py -k "aiter"

• Before: test_modular_kernel_combinations_multigpu[...] processes died with the AttributeError above (see failure traceback in PR context).
• After: 54 passed, 6 skipped, 21 deselected (skips are pre-existing/expected — mxfp4 unsupported here, and known AITER×DeepEP-HT illegal-memory-access cases tracked in [Bug][ROCm] AiterExperts + DeepEPHT/Mori PrepareAndFinalize: illegal memory access at world_size=2 (gfx942) #57029).

Also spot-checked individually:

• test_aiter_moe_token_padding_garbage_rows[single_nan-unquantized] — PASSED
• test_modular_kernel_combinations_multigpu[2048-1024-32-dtype1-None-MoEActivation.SILU-DeepEPHTPrepareAndFinalize-AiterExperts-2] — PASSED (previously crashed)

This change doesn't affect model output/accuracy, so no model eval was run.

🤖 Generated with Claude Code

[divakar-amd]

Moved these fixes to #60322. Hence, Closing this draft PR