[AMD][AgentX] Kimi-K3 MXFP4 MI355X AgentX, DSpark only / Kimi-K3 MXFP4 MI355X AgentX 仅 DSpark 配置 - #2403

Open

hyukjlee wants to merge 56 commits into

mainfrom

kimik3-agentic-smoke

Commits on Jul 30, 2026

[AMD][AgentX] kimik3: MXFP4 MI355X agentic recipe from the upstream vLLM recipe

hyukjlee
and
claude
committed
8d026ee
[AMD][AgentX] kimik3: set agentic dram-utilization to 0.80

hyukjlee
and
claude
committed
8d9aa88
[AMD][AgentX] kimik3: LMCache requires prefix caching on K3 -- force it

hyukjlee
and
claude
committed
47a6e68
[AMD][AgentX] kimik3: cap LMCache L1 to 512 GB (eager-allocation ceiling)

hyukjlee
and
claude
committed
82c3a93
[AMD][AgentX] kimik3: add a k2.7 LMCache server profile, drop the 512 GB ceiling

hyukjlee
and
claude
committed
58bb118
[AMD][AgentX] kimik3: LMCache forces --max-num-batched-tokens 768 (Mamba block constraint)

hyukjlee
and
claude
committed
911db9e
[AMD][AgentX] kimik3: pin LMCache chunk-size to a multiple of block_size

hyukjlee
and
claude
committed
0c5a952
[AMD][AgentX] kimik3: select the LMCache server profile from config

hyukjlee
and
claude
committed
73c9c68
[AMD][AgentX] kimik3: enable prefix caching by default on every arm

hyukjlee
and
claude
committed
592a13b
[AMD][AgentX] kimik3: bringup key = none + both LMCache server profiles at c8

hyukjlee
and
claude
committed
6775c12
[AMD][AgentX] mi355x-amds: exclude mia1-p01-g14 from salloc

hyukjlee
and
claude
committed
f55e422
[AMD][AgentX] mi355x-amds: deny g15/g18 too; add kimik3 retry key

hyukjlee
and
claude
committed
5cecde9
[AMD][AgentX] kimik3: add the agentic EVAL_ONLY path (guidance backend + 64-way)

hyukjlee
and
claude
committed
37f0209
[AMD][AgentX] kimik3: use the reference LMCache (pip lmcache==0.5.1, L1 512, chunk 1024, mnbt 4096)

hyukjlee
and
claude
committed
d811a76
[AMD][AgentX] kimik3: add lmcache-budget (L1 from the DRAM budget, not the flat 512)

hyukjlee
and
claude
committed
471c4d6
[AMD][AgentX] kimik3: gpu-memory-utilization 0.95 -> 0.88

hyukjlee
and
claude
committed
09a6591
[AMD][AgentX] kimik3: 0.5.1 enforces the Mamba-hybrid limits too -- re-pin 768

hyukjlee
and
claude
committed
f146915
[AMD][AgentX] kimik3: add native KV-offload tiers (vllm-simple, vllm-native)

hyukjlee
and
claude
committed
218117c
[AMD][AgentX] kimik3: full 3600s sweep -- none c[1,2,4] + vllm-simple c[4,8,16,32]

hyukjlee
and
claude
committed
f27aa5f
[AMD][AgentX] kimik3: denylist g11 (GPU faults), retry none c[1,2,4]

hyukjlee
and
claude
committed
2818737
[AMD][AgentX] kimik3: add fp8 KV-cache probe (vllm-simple-fp8, c8)

hyukjlee
and
claude
committed
d891f75
[AMD][AgentX] kimik3: don't enable prefix caching on the no-offload arm

hyukjlee
and
claude
committed
ff2eaad
[AMD][AgentX] kimik3: fp8 KV on a b128-tuned MLA image; soften the prefix-caching claim

hyukjlee
and
claude
committed
347cdb7
[AMD][AgentX] mi355x-amds: turn the salloc denylist off by default

hyukjlee
and
claude
committed
eeb9d1a
[AMD][AgentX] kimik3: patch aiter Gluon MLA kernel from gist before serving

authored and
hyukjlee
committed
c62b938
Revert "[AMD][AgentX] kimik3: don't enable prefix caching on the no-offload arm"

hyukjlee
and
claude
committed
845a0ec
[AMD][AgentX] kimik3: widen worker RPC timeout for long-context fp8-KV arm

authored and
hyukjlee
committed
a15e7d5
[AMD][AgentX] kimik3: add bf16-KV offload isolation arm; disable RPC-timeout+mla_gluon patches

authored and
hyukjlee
committed
e193dbb
[AMD][AgentX] kimik3: point the fp8-KV cell at the fp8-kv-fixed image

hyukjlee
and
claude
committed
8334611
[AMD][AgentX] mi355x-amds: exclude mia1-p01-g11 from salloc by default

hyukjlee
and
claude
committed
36b43db
[AMD][AgentX] kimik3: add fp8kvtest arm (fp8 KV, no offload) for isolation

authored and
hyukjlee
committed
430e8cf
[AMD][AgentX] kimik3: disable forced fp8 KV hardcode

authored and
hyukjlee
committed
e17f084
[AMD][AgentX] kimik3: pin the mla_gluon gist, point fp8kv at the proven stack

hyukjlee
and
claude
committed
606584c
[AMD][AgentX] kimik3: make fp8 KV the default and add the full fp8 sweep

hyukjlee
and
claude
committed
1f7ce51
[AMD][AgentX] kimik3: fix fp8kvtest search-space, it was failing validation

hyukjlee
and
claude
committed
124599f
[AMD][AgentX] kimik3: add fp8retry arm (none c[1,16] + vllm-simple-fp8 c8)

hyukjlee
and
claude
committed
a5dd587
feat(amd): add Kimi-K3 DSpark AgentX variant

edwingao28
authored and
hyukjlee
committed
c192d17
[AMD][AgentX] kimik3: pin the KV page to 64 on an fp8+offload arm

hyukjlee
and
claude
committed
d78983b
[AMD][AgentX] kimik3: force lm-eval and add an fp8 kvnone c8 eval arm

hyukjlee
and
claude
committed
f495ab7
[AMD][AgentX] kimik3: try lazy offload for fp8 KV; --block-size is inert on K3

hyukjlee
and
claude
committed
b9f5fc6
fix(amd): align Kimi-K3 DSpark with ROCm baseline

edwingao28
authored and
hyukjlee
committed
bd7678a
[AMD][AgentX] kimik3: pair eager+lazy fp8 offload on the new AIPerf warmup

hyukjlee
and
claude
committed
220e0ee
[AMD][AgentX] kimik3: fp8-KV no-offload curve at GMU 0.90, conc 1/4/8/16

hyukjlee
and
claude
committed
0af1730
fix(amd): reduce Kimi-K3 DSpark sequence slots

edwingao28
authored and
hyukjlee
committed
f2feeb8
[AMD][AgentX] kimik3: revert GMU to 0.88, rename the fp8 no-offload key

hyukjlee
and
claude
committed
86a097b
[AMD][AgentX] kimik3: enable LMCache via upstream's validated K3 support

hyukjlee
and
claude
committed
57ca399
[AMD][AgentX] kimik3: give the LMCache arm headroom (mns 32, GMU 0.85)

hyukjlee
and
claude
committed
938a922
[AMD][AgentX] kimik3: size the LMCache L1 pool from the DRAM budget

hyukjlee
and
claude
committed
8eaf651
[AMD][AgentX] kimik3: L1=512GB — 1360GB breaks GPU IPC registration

hyukjlee
and
claude
committed
f6f14be
[AMD][AgentX] kimik3: add c16 to the LMCache arm

hyukjlee
and
claude
committed
463f1a9

Commits on Jul 31, 2026

Commits on Aug 1, 2026

Merge branch 'main' into kimik3-agentic-smoke
seungrokj
authored
a4c104a

Commits on Aug 2, 2026

[AMD][AgentX] kimik3: reduce the PR to the DSpark arm only

functionstackx
and
claude
committed
0b50bea