- Register DeepEPV2PrepareAndFinalize and TrtLlmFp8ExpertsModular in
modular kernel test framework (mk_objects.py)
- Add model_config to test VllmConfig so all2all_utils can read
enforce_eager for cudagraph detection
- Add weight conversion for TrtLLM BlockMajorK format in test framework
- Separate DeepEP v1/v2 dependency checks in test validation
- Allow TrtLlmFp8ExpertsModular with DeepEP v2 parallel config
- Auto-select TrtLLM backend for DeepEP v2 contiguous layout on Blackwell
- Remove unnecessary torch.where sanitization in DeepEP v2 decode path
- Replace torch.cat with pre-allocated tensor in prefill path
- Always create ExpertTokensMetadata in decode mode
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
