[AMD][AgentX] DSv4 MI355X agentic disagg: add high-concurrency dots using DEP8 - #2308

Open

ichbinblau wants to merge 22 commits into

mainfrom

amd/agentx-v1.0-th-hicon-dpa

Commits on Jul 29, 2026

HiCache sizing fixes (ratio vs size), NCCL/image updates, disable prefill cudagraph for DSv4 agentic disagg

ichbinblau
and
cursoragent
committed
Update DPA config, add RDMA/GPU pre-flight gates to job.slurm, and fix router policy env var/name

ichbinblau
and
cursoragent
committed
Add perf-changelog.yaml entry for dsv4-fp4-mi355x-sglang-disagg-agentic-hicache

ichbinblau
and
cursoragent
committed
remove disable_cuda_graph: false setting from decode
ichbinblau
committed
add env PREFILL_DP_ATTN and DECODE_DP_ATTN
ichbinblau
committed
fix(dsv4): forward PREFILL_EP/DECODE_EP/*_DP_ATTN into agentic server container

ichbinblau
and
cursoragent
committed
feat(agentic): add multi-node agentic (SWE-bench) eval dispatch support

ichbinblau
and
cursoragent
committed
Mark dsv4-fp4-mi355x-sglang-disagg-agentic-hicache changelog entry evals-only

ichbinblau
and
cursoragent
committed
Fix multinode agentic eval-only CI: forward Modal creds and stage eval artifacts reliably.

ichbinblau
and
cursoragent
committed
Harden multinode agentic Modal credential setup.

ichbinblau
and
cursoragent
committed
Switch multinode agentic eval to GSM8K lm-eval with correct metadata.

ichbinblau
and
cursoragent
committed
Harden multinode agentic eval metadata and GPU pre-flight for DEP8 CI.

ichbinblau
and
cursoragent
committed
Add EP1 GSM8K eval changelog entry alongside DEP8 c128 eval.

ichbinblau
and
cursoragent
committed
Revert "Add EP1 GSM8K eval changelog entry alongside DEP8 c128 eval."
ichbinblau
committed
include node mia1-p01-g10 and mia1-p01-g16
ichbinblau
committed
exclude mia1-p01-g10
ichbinblau
committed
add con=1 in config
ichbinblau
committed
remove con=64
ichbinblau
committed
remove SGLANG_AITER_MLA_PERSIST
ichbinblau
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[AMD][AgentX] DSv4 MI355X agentic disagg: add high-concurrency dots using DEP8 - #2308

[AMD][AgentX] DSv4 MI355X agentic disagg: add high-concurrency dots using DEP8#2308
ichbinblau wants to merge 22 commits into
mainfrom
amd/agentx-v1.0-th-hicon-dpa

Commits on Jul 29, 2026

HiCache sizing fixes (ratio vs size), NCCL/image updates, disable prefill cudagraph for DSv4 agentic disagg

Update DPA config, add RDMA/GPU pre-flight gates to job.slurm, and fix router policy env var/name

Add perf-changelog.yaml entry for dsv4-fp4-mi355x-sglang-disagg-agentic-hicache

remove disable_cuda_graph: false setting from decode

add env PREFILL_DP_ATTN and DECODE_DP_ATTN

fix(dsv4): forward PREFILL_EP/DECODE_EP/*_DP_ATTN into agentic server container

feat(agentic): add multi-node agentic (SWE-bench) eval dispatch support

Mark dsv4-fp4-mi355x-sglang-disagg-agentic-hicache changelog entry evals-only

Fix multinode agentic eval-only CI: forward Modal creds and stage eval artifacts reliably.

Harden multinode agentic Modal credential setup.

Switch multinode agentic eval to GSM8K lm-eval with correct metadata.

Harden multinode agentic eval metadata and GPU pre-flight for DEP8 CI.

Add EP1 GSM8K eval changelog entry alongside DEP8 c128 eval.

Revert "Add EP1 GSM8K eval changelog entry alongside DEP8 c128 eval."

include node mia1-p01-g10 and mia1-p01-g16

exclude mia1-p01-g10

add con=1 in config

remove con=64

remove SGLANG_AITER_MLA_PERSIST

Commits on Jul 30, 2026

Merge branch 'main' into amd/agentx-v1.0-th-hicon-dpa

fix(amd): extend GPU drain poll, trim node excludes, pin aiperf

fix(agentic): use explicit Python 3.11 for isolated aiperf venv