[AMD][AgentX] DeepSeek-V4 MI355X agentic disaggregated benchmarking#2170

Open

ichbinblau wants to merge 40 commits into

mainfrom

amd/agentx-v1.0-yanfei

Commits on Jul 14, 2026

[AMD][AgentX] Add DeepSeek V4 MI355X agentic disagg

cquil11
authored and
ichbinblau
committed
run-sweep: fix broken conc-list interpolation for multi-node agentic

ichbinblau
and
cursoragent
committed
bump image to rocm/sgl-dev:sglang-0.5.14-rocm720-mi35x-mori-0706
ichbinblau
committed
run-sweep: fix sequence-typed conc input for multi-node agentic

ichbinblau
and
cursoragent
committed
agentic: run one task per concurrency for multi-node sweeps

ichbinblau
and
cursoragent
committed
Bump image to latest upstream image with 4 PR fixes included.
ichbinblau
committed
update con-list
ichbinblau
committed
setup_deps: remove disabled decode_tp_queue_agree patch

ichbinblau
and
cursoragent
committed
setup_deps: broaden GLM transformers gate; disable prefill bootstrap-desync patch

ichbinblau
and
cursoragent
committed
agentic: node-0 sibling benchmark-client container for DSv4 sweeps (#2147)

AMD-yanfeiwang
authored and
ichbinblau
committed
enable log info and extend timeout
ichbinblau
committed
fix: address PR 2170 review comments

Duyi-Wang
authored and
ichbinblau
committed
setup_deps: remove prefill bootstrap-desync patch entirely

ichbinblau
and
cursoragent
committed
fix perf changelog

billishyahao
authored and
ichbinblau
committed
consolidate sglang agentic envs
ichbinblau
committed
select models.yaml recipe by IS_AGENTIC (AgentX/DI variants)

ichbinblau
and
cursoragent
committed
DPTP enabling.
ichbinblau
committed
tune hicache ratio, router barrier timeout, and agentic conc sweep

ichbinblau
and
cursoragent
committed
raise hicache ratio to 16
ichbinblau
committed
fix: restore qwen3.5 changelog entry removed during rebase

ichbinblau
and
cursoragent
committed
fix: validate agentic hicache master config for changelog gate

ichbinblau
and
cursoragent
committed
feat: wire KV_P2P_TRANSFER env into SGLang launch flags

ichbinblau
and
cursoragent
committed
chore: drop max_total_tokens from DSv4 AgentX prefill no_dp recipe

ichbinblau
and
cursoragent
committed
tune: raise router barrier timeout to 3000s for worker nodes

ichbinblau
and
cursoragent
committed
test: fix sweep config fixtures for kv-p2p-transfer validation

ichbinblau
and
cursoragent
committed
refactor: centralize sync barrier timeouts for DSV4 startup

ichbinblau
and
cursoragent
committed
fix: forward KV offload backend metadata to agentic client

ichbinblau
and
cursoragent
committed
Merge branch 'main' into amd/agentx-v1.0-yanfei
ichbinblau
authored
revert: drop metadata-derived KV_OFFLOAD_BACKEND from DSv4 recipe

ichbinblau
and
cursoragent
committed
fix: quote KV_OFFLOAD_BACKEND_METADATA in docker run env forwarding

ichbinblau
and
cursoragent
committed
Merge remote-tracking branch 'origin/amd/agentx-v1.0-yanfei'
ichbinblau
committed
Fix client.env generation to flatten multi-line env values

ichbinblau
and
cursoragent
committed
Re-serialize KV_OFFLOAD_BACKEND_METADATA as compact JSON instead of stripping newlines

ichbinblau
and
cursoragent
committed
Fix regression: skip JSON round-trip for empty KV_OFFLOAD_BACKEND_METADATA

ichbinblau
and
cursoragent
committed
Merge remote-tracking branch 'origin/main' into agent/batch-null-2170
cquil11
committed
fix(changelog): append PR entry after main
cquil11
committed
fix(changelog): omit null optional metadata
cquil11
committed
fix(agentx): set DeepSeek reasoning effort high
cquil11
committed
Merge remote-tracking branch 'origin/main' into amd/agentx-v1.0-yanfei

ichbinblau
and
cursoragent
committed

Commits on Jul 15, 2026

Remove stale patches/README.md documenting an already-retired mori_conn.py overlay

ichbinblau
and
cursoragent
committed