Skip to content

Add DSR1-0528 FP8 MI300X vLLM PD-disaggregation#1

Closed
edwinlim0919 wants to merge 1 commit into
mainfrom
dsr1-fp8-mi300x-vllm-disagg
Closed

Add DSR1-0528 FP8 MI300X vLLM PD-disaggregation#1
edwinlim0919 wants to merge 1 commit into
mainfrom
dsr1-fp8-mi300x-vllm-disagg

Conversation

@edwinlim0919

Copy link
Copy Markdown
Collaborator

Adds the dsr1-fp8-mi300x-vllm config for DeepSeek-R1-0528 FP8 on AMD MI300X using vLLM PD-disaggregation (MoRIIO + vllm-router).

Topology: 2P2D TP8 with MTP3 speculative decoding.
Image: vllm-mori-pd:milestone4-aiterwheel
Runner: mi300x-disagg

Files changed:

  • .github/configs/amd-master.yaml: new config entry
  • benchmarks/multi_node/amd_utils/models_vllm.yaml: DeepSeek-R1-0528 model flags/env
  • benchmarks/multi_node/dsr1_fp8_mi300x_vllm-disagg.sh: recipe
  • runners/launch_mi300x-amds.sh: IS_MULTINODE branch
  • benchmarks/multi_node/amd_utils/{env.sh, server_vllm.sh, job.slurm, bench.sh}: Thor/bnxt fabric adaptations, MoRIIO extra config, chat-template for MTP
  • perf-changelog.yaml: entry for dsr1-fp8-mi300x-vllm

@edwinlim0919 edwinlim0919 force-pushed the dsr1-fp8-mi300x-vllm-disagg branch 3 times, most recently from 62ed295 to 5ad2a82 Compare June 24, 2026 06:27
- .github/configs/amd-master.yaml: add dsr1-fp8-mi300x-vllm (2P2D TP8 MTP3)
- benchmarks/multi_node/amd_utils/models_vllm.yaml: add DeepSeek-R1-0528 entry with M4 flags
- benchmarks/multi_node/dsr1_fp8_mi300x_vllm-disagg.sh: new disagg recipe
- runners/launch_mi300x-amds.sh: add IS_MULTINODE branch for MI300X disagg
- benchmarks/multi_node/amd_utils/env.sh: Thor/bnxt explicit-match NCCL_IB_HCA + MORI_RDMA_DEVICES
- benchmarks/multi_node/amd_utils/server_vllm.sh: 172.29.x RDMA IP, MoRIIO extra config, IS_MTP
- benchmarks/multi_node/amd_utils/job.slurm: forward RDMA env vars, no host lib mounts for vllm
- benchmarks/multi_node/amd_utils/bench.sh: --use-chat-template for vllm-disagg MTP
- perf-changelog.yaml: document new config

Image: vllm-mori-pd:milestone4-aiterwheel
@edwinlim0919 edwinlim0919 force-pushed the dsr1-fp8-mi300x-vllm-disagg branch from 5ad2a82 to 71e7ad1 Compare June 24, 2026 07:07
@edwinlim0919 edwinlim0919 deleted the dsr1-fp8-mi300x-vllm-disagg branch June 24, 2026 07:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant