Repository navigation
[ROCm][CI] Mirror generic GEMM-RS/AR on MI355 - #58433
Conversation
Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Andreas Karatzas <andreas.karatzas@protonmail.com>
|
/ci run |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
✅ Triggered Buildkite CI #90750 for commit |
AndreasKaratzas
left a comment
AndreasKaratzas
left a comment
There was a problem hiding this comment.
https://buildkite.com/vllm/ci/builds/90750/table?jid=01a0cfdb-cbbb-4bba-b8e3-7a88e7352ea4&tab=output
Observing an intermittent failure here. Investigating.
|
Setting it to draft mode until ROCm/aiter#5807 is upstreamed. |
|
/ci run |
|
✅ Triggered Buildkite CI #91536 for commit |
|
/ci run |
|
✅ CI is already running for this commit: https://buildkite.com/vllm/ci/builds/91536 |
Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Andreas Karatzas <andreas.karatzas@protonmail.com>
|
/ci run |
|
✅ Triggered Buildkite CI #91782 for commit |
The AMD pipeline derives block step keys from labels and cannot upload keys containing a period. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Andreas Karatzas <andreas.karatzas@protonmail.com>
|
/ci run |
|
✅ Triggered Buildkite CI #91790 for commit |
|
/ci run |
|
❌ This PR is 5 commits behind upstream |
|
/ci run |
|
✅ Triggered Buildkite CI #91818 for commit |
3 participants
gemm-rs-ar-2xb200parent and a matching native AMD job, with four GPUs, a 20-minute timeout, and source dependencies covering the tested production paths.test_rocm_gemm_rs_ar.pyfor BF16 and native MXFP8 at TP2 and TP4, exercising productionRowParallelLinear, AITER all-reduce, and RCCL reduce-scatter with explicit backend assertions.The generic GEMM-RS/AR parent introduced by #57428 extends the former Kimi-specific coverage with MXFP8 and additional shape and padding checks. The older #50519 proposal uses the removed Kimi helper and needs its test workload updated along with its parent association. This change adds coverage for the expanded contract through supported ROCm execution paths and checks native MXFP8 dispatch explicitly. It establishes correctness coverage; fused GEMM/communication overlap and performance parity remain unmeasured.
AI assistance was used for implementation, review, local validation, and this description.