perf(agentx): refresh K3 MI355X vLLM recipe with DCP8 MTP arm - #2810

Open

seungrokj wants to merge 25 commits into

mainfrom

amd/agentx_k3_vllm_0903_2

Commits on Sep 3, 2026

perf(agentx): refresh K3 MI355X vLLM recipe with DCP8 MTP arm

seungrokj
and
claude
committed
71acdd4
docs(changelog): link K3 MI355X vLLM DCP8 MTP refresh PR

seungrokj
and
claude
committed
c281dbe
style(agentx): add trailing newline to K3 MI355X benchmark script

seungrokj
and
claude
committed
7b93468
perf(agentx): drop c52 from K3 MI355X DCP8 arm

seungrokj
and
claude
committed
db7732d
perf(agentx): gate K3 LMCache GPU workers on DCP; trim sweep to c14/c48

seungrokj
and
claude
committed
0a32492
perf(agentx): lower K3 c12/c14 GMU to 0.88; run c14-only smoke test

seungrokj
and
claude
committed
1701d42
perf(agentx): run K3 c12/c14 GMU 0.88 smoke test

seungrokj
and
claude
committed
aefefb3
perf(agentx): lower K3 c12/c14 GMU to 0.85

seungrokj
and
claude
committed
5cebe65
perf(agentx): test K3 c14 GPU-resident (no KV offload), GMU 0.9

seungrokj
and
claude
committed
1b24693
perf(agentx): bump K3 image to nightly-73029d42; c14 no-spec GPU-resident

seungrokj
and
claude
committed
ffc4d74
perf(agentx): run K3 c48 DCP8 LMCache smoke test on nightly-73029d42

seungrokj
and
claude
committed
99d0671
perf(agentx): run K3 c14 GPU-resident with MTP spec on nightly-73029d42

seungrokj
and
claude
committed
34d93c9
perf(agentx): drop K3 high-conc max-num-batched-tokens to 4096 (OOM fix)

seungrokj
and
claude
committed
883d6e3
perf(agentx): activate K3 c48 DCP8 arm for OOM-fix validation

seungrokj
and
claude
committed
60acfb3
perf(agentx): gate K3 LMCache chunk-size on DCP; pin dev60 wheel; run c14

seungrokj
and
claude
committed
47745bf
perf(agentx): gate K3 LMCache wheel on DCP; run c14 on dev60

seungrokj
and
claude
committed
f48e07a
fix(agentx): define LMCACHE_RELEASE for the K3 index URL

seungrokj
and
claude
committed
ed613b6
perf(agentx): pin K3 LMCache to v0.5.5rc4 from the release asset page

seungrokj
and
claude
committed
bbce1d2
perf(agentx): pin K3 LMCache to dev89; run c14 (dcp1) and c48 (dcp8)

seungrokj
and
claude
committed
0a3b572
perf(agentx): revert K3 image to 7c5dc571; c48 label to dev89

seungrokj
and
claude
committed
d5e7a1c
perf(agentx): bump K3 image to 73029d42; run full conc sweep on dev89

seungrokj
and
claude
committed
05c3f3f
Merge branch 'main' into amd/agentx_k3_vllm_0903_2
seungrokj
authored
de3cc42
perf(agentx): add conc 52 to K3 dcp8 sweep

seungrokj
and
claude
committed
932280a
perf(agentx): revert K3 image to 7c5dc571

seungrokj
and
claude
committed
d9f6855
fix(ci): pin pydantic>=2 for perf-changelog validation

seungrokj
and
claude
committed
cbf74f5

Navigation Menu

Uh oh!

perf(agentx): refresh K3 MI355X vLLM recipe with DCP8 MTP arm - #2810

perf(agentx): refresh K3 MI355X vLLM recipe with DCP8 MTP arm#2810
seungrokj wants to merge 25 commits into
mainfrom
amd/agentx_k3_vllm_0903_2

Commits on Sep 3, 2026

perf(agentx): refresh K3 MI355X vLLM recipe with DCP8 MTP arm

docs(changelog): link K3 MI355X vLLM DCP8 MTP refresh PR

style(agentx): add trailing newline to K3 MI355X benchmark script

perf(agentx): drop c52 from K3 MI355X DCP8 arm

perf(agentx): gate K3 LMCache GPU workers on DCP; trim sweep to c14/c48

perf(agentx): lower K3 c12/c14 GMU to 0.88; run c14-only smoke test

perf(agentx): run K3 c12/c14 GMU 0.88 smoke test

perf(agentx): lower K3 c12/c14 GMU to 0.85

perf(agentx): test K3 c14 GPU-resident (no KV offload), GMU 0.9

perf(agentx): bump K3 image to nightly-73029d42; c14 no-spec GPU-resident

perf(agentx): run K3 c48 DCP8 LMCache smoke test on nightly-73029d42

perf(agentx): run K3 c14 GPU-resident with MTP spec on nightly-73029d42

perf(agentx): drop K3 high-conc max-num-batched-tokens to 4096 (OOM fix)

perf(agentx): activate K3 c48 DCP8 arm for OOM-fix validation

perf(agentx): gate K3 LMCache chunk-size on DCP; pin dev60 wheel; run c14

perf(agentx): gate K3 LMCache wheel on DCP; run c14 on dev60

fix(agentx): define LMCACHE_RELEASE for the K3 index URL

perf(agentx): pin K3 LMCache to v0.5.5rc4 from the release asset page

perf(agentx): pin K3 LMCache to dev89; run c14 (dcp1) and c48 (dcp8)

perf(agentx): revert K3 image to 7c5dc571; c48 label to dev89

perf(agentx): bump K3 image to 73029d42; run full conc sweep on dev89

Merge branch 'main' into amd/agentx_k3_vllm_0903_2

perf(agentx): add conc 52 to K3 dcp8 sweep

perf(agentx): revert K3 image to 7c5dc571

fix(ci): pin pydantic>=2 for perf-changelog validation