[AMD][MI355X] Add DSv4 FP4 agentic MTP recipe with DPA tuning#2254

Open

seungrokj wants to merge 16 commits into

mainfrom

amd/agentx_dsv4_sgl_mtp_0717

Commits on Jul 19, 2026

[AMD][MI355X] Add DSv4 FP4 agentic MTP recipe with DPA tuning

seungrokj
and
claude
committed
fix(changelog): add pr-link for dsv4-fp4-mi355x-sglang-agentic-mtp entry

seungrokj
and
claude
committed
fix(changelog): trim verbose description for agentic-mtp entry

seungrokj
and
claude
committed
fix(config): remove duplicate dsv4-fp4-mi355x-sglang-agentic-hicache entry

seungrokj
and
claude
committed
fix(validation): add run-eval/eval-only fields to agentic matrix entries

seungrokj
and
claude
committed
fix(agentic-mtp): tune DPA memory, prefill chunk, and batch sizing

seungrokj
and
claude
committed
fix(agentic-mtp): revert mem_fraction_static override, narrow sweep to c48

seungrokj
and
claude
committed
fix(agentic-mtp): stop dividing CUDA_GRAPH_MAX_BS by TP for DPA

seungrokj
and
claude
committed
fix(agentic-mtp): restore MAX_RUNNING_REQUESTS to 2*CONC

seungrokj
and
claude
committed
fix(agentic-mtp): reduce chunked prefill back to 8K per scheduler

seungrokj
and
claude
committed
fix(agentic-mtp): restore GPU_MAX_HW_QUEUES=5 for ROCm overlap

seungrokj
and
claude
committed
fix(agentic-mtp): clean up script to match ref, expand sweep concurrencies

seungrokj
and
claude
committed
fix(validation): remove run_eval/eval_only from MultiNodeAgenticMatrixEntry

seungrokj
and
claude
committed
fix(agentic-mtp): add MAX_RUNNING_REQUESTS=2*CONC with comment

seungrokj
and
claude
committed
fix(agentic-mtp): drop c52 from sweep to avoid GPU OOM

seungrokj
and
claude
committed
feat(agentic-mtp): add cache metrics capture and eval-only support

seungrokj
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[AMD][MI355X] Add DSv4 FP4 agentic MTP recipe with DPA tuning#2254

[AMD][MI355X] Add DSv4 FP4 agentic MTP recipe with DPA tuning#2254
seungrokj wants to merge 16 commits into
mainfrom
amd/agentx_dsv4_sgl_mtp_0717

Commits on Jul 19, 2026

[AMD][MI355X] Add DSv4 FP4 agentic MTP recipe with DPA tuning

fix(changelog): add pr-link for dsv4-fp4-mi355x-sglang-agentic-mtp entry

fix(changelog): trim verbose description for agentic-mtp entry

fix(config): remove duplicate dsv4-fp4-mi355x-sglang-agentic-hicache entry

fix(validation): add run-eval/eval-only fields to agentic matrix entries

fix(agentic-mtp): tune DPA memory, prefill chunk, and batch sizing

fix(agentic-mtp): revert mem_fraction_static override, narrow sweep to c48

fix(agentic-mtp): stop dividing CUDA_GRAPH_MAX_BS by TP for DPA

fix(agentic-mtp): restore MAX_RUNNING_REQUESTS to 2*CONC

fix(agentic-mtp): reduce chunked prefill back to 8K per scheduler

fix(agentic-mtp): restore GPU_MAX_HW_QUEUES=5 for ROCm overlap

fix(agentic-mtp): clean up script to match ref, expand sweep concurrencies

fix(validation): remove run_eval/eval_only from MultiNodeAgenticMatrixEntry

fix(agentic-mtp): add MAX_RUNNING_REQUESTS=2*CONC with comment

fix(agentic-mtp): drop c52 from sweep to avoid GPU OOM

feat(agentic-mtp): add cache metrics capture and eval-only support