Skip to content
Open
Changes from all commits
Commits
Show all changes
57 commits
Select commit Hold shift + click to select a range
8d026ee
[AMD][AgentX] kimik3: MXFP4 MI355X agentic recipe from the upstream v…
hyukjlee Jul 28, 2026
8d9aa88
[AMD][AgentX] kimik3: set agentic dram-utilization to 0.80
hyukjlee Jul 28, 2026
47a6e68
[AMD][AgentX] kimik3: LMCache requires prefix caching on K3 -- force it
hyukjlee Jul 28, 2026
82c3a93
[AMD][AgentX] kimik3: cap LMCache L1 to 512 GB (eager-allocation ceil…
hyukjlee Jul 28, 2026
58bb118
[AMD][AgentX] kimik3: add a k2.7 LMCache server profile, drop the 512…
hyukjlee Jul 28, 2026
911db9e
[AMD][AgentX] kimik3: LMCache forces --max-num-batched-tokens 768 (Ma…
hyukjlee Jul 28, 2026
0c5a952
[AMD][AgentX] kimik3: pin LMCache chunk-size to a multiple of block_size
hyukjlee Jul 28, 2026
73c9c68
[AMD][AgentX] kimik3: select the LMCache server profile from config
hyukjlee Jul 28, 2026
592a13b
[AMD][AgentX] kimik3: enable prefix caching by default on every arm
hyukjlee Jul 28, 2026
6775c12
[AMD][AgentX] kimik3: bringup key = none + both LMCache server profil…
hyukjlee Jul 28, 2026
f55e422
[AMD][AgentX] mi355x-amds: exclude mia1-p01-g14 from salloc
hyukjlee Jul 28, 2026
5cecde9
[AMD][AgentX] mi355x-amds: deny g15/g18 too; add kimik3 retry key
hyukjlee Jul 28, 2026
37f0209
[AMD][AgentX] kimik3: add the agentic EVAL_ONLY path (guidance backen…
hyukjlee Jul 28, 2026
d811a76
[AMD][AgentX] kimik3: use the reference LMCache (pip lmcache==0.5.1, …
hyukjlee Jul 28, 2026
471c4d6
[AMD][AgentX] kimik3: add lmcache-budget (L1 from the DRAM budget, no…
hyukjlee Jul 28, 2026
09a6591
[AMD][AgentX] kimik3: gpu-memory-utilization 0.95 -> 0.88
hyukjlee Jul 28, 2026
f146915
[AMD][AgentX] kimik3: 0.5.1 enforces the Mamba-hybrid limits too -- r…
hyukjlee Jul 28, 2026
218117c
[AMD][AgentX] kimik3: add native KV-offload tiers (vllm-simple, vllm-…
hyukjlee Jul 28, 2026
f27aa5f
[AMD][AgentX] kimik3: full 3600s sweep -- none c[1,2,4] + vllm-simple…
hyukjlee Jul 28, 2026
2818737
[AMD][AgentX] kimik3: denylist g11 (GPU faults), retry none c[1,2,4]
hyukjlee Jul 29, 2026
d891f75
[AMD][AgentX] kimik3: add fp8 KV-cache probe (vllm-simple-fp8, c8)
hyukjlee Jul 29, 2026
ff2eaad
[AMD][AgentX] kimik3: don't enable prefix caching on the no-offload arm
hyukjlee Jul 29, 2026
347cdb7
[AMD][AgentX] kimik3: fp8 KV on a b128-tuned MLA image; soften the pr…
hyukjlee Jul 29, 2026
eeb9d1a
[AMD][AgentX] mi355x-amds: turn the salloc denylist off by default
hyukjlee Jul 29, 2026
c62b938
[AMD][AgentX] kimik3: patch aiter Gluon MLA kernel from gist before s…
seungrokj Jul 29, 2026
845a0ec
Revert "[AMD][AgentX] kimik3: don't enable prefix caching on the no-o…
hyukjlee Jul 29, 2026
a15e7d5
[AMD][AgentX] kimik3: widen worker RPC timeout for long-context fp8-K…
seungrokj Jul 29, 2026
e193dbb
[AMD][AgentX] kimik3: add bf16-KV offload isolation arm; disable RPC-…
seungrokj Jul 29, 2026
8334611
[AMD][AgentX] kimik3: point the fp8-KV cell at the fp8-kv-fixed image
hyukjlee Jul 29, 2026
36b43db
[AMD][AgentX] mi355x-amds: exclude mia1-p01-g11 from salloc by default
hyukjlee Jul 29, 2026
430e8cf
[AMD][AgentX] kimik3: add fp8kvtest arm (fp8 KV, no offload) for isol…
seungrokj Jul 29, 2026
e17f084
[AMD][AgentX] kimik3: disable forced fp8 KV hardcode
seungrokj Jul 29, 2026
606584c
[AMD][AgentX] kimik3: pin the mla_gluon gist, point fp8kv at the prov…
hyukjlee Jul 29, 2026
1f7ce51
[AMD][AgentX] kimik3: make fp8 KV the default and add the full fp8 sweep
hyukjlee Jul 29, 2026
124599f
[AMD][AgentX] kimik3: fix fp8kvtest search-space, it was failing vali…
hyukjlee Jul 29, 2026
a5dd587
[AMD][AgentX] kimik3: add fp8retry arm (none c[1,16] + vllm-simple-fp…
hyukjlee Jul 29, 2026
c192d17
feat(amd): add Kimi-K3 DSpark AgentX variant
edwingao28 Jul 29, 2026
d78983b
[AMD][AgentX] kimik3: pin the KV page to 64 on an fp8+offload arm
hyukjlee Jul 30, 2026
f495ab7
[AMD][AgentX] kimik3: force lm-eval and add an fp8 kvnone c8 eval arm
hyukjlee Jul 30, 2026
b9f5fc6
[AMD][AgentX] kimik3: try lazy offload for fp8 KV; --block-size is in…
hyukjlee Jul 30, 2026
bd7678a
fix(amd): align Kimi-K3 DSpark with ROCm baseline
edwingao28 Jul 30, 2026
220e0ee
[AMD][AgentX] kimik3: pair eager+lazy fp8 offload on the new AIPerf w…
hyukjlee Jul 30, 2026
0af1730
[AMD][AgentX] kimik3: fp8-KV no-offload curve at GMU 0.90, conc 1/4/8/16
hyukjlee Jul 30, 2026
f2feeb8
fix(amd): reduce Kimi-K3 DSpark sequence slots
edwingao28 Jul 30, 2026
86a097b
[AMD][AgentX] kimik3: revert GMU to 0.88, rename the fp8 no-offload key
hyukjlee Jul 30, 2026
57ca399
[AMD][AgentX] kimik3: enable LMCache via upstream's validated K3 support
hyukjlee Jul 30, 2026
938a922
[AMD][AgentX] kimik3: give the LMCache arm headroom (mns 32, GMU 0.85)
hyukjlee Jul 30, 2026
8eaf651
[AMD][AgentX] kimik3: size the LMCache L1 pool from the DRAM budget
hyukjlee Jul 30, 2026
f6f14be
[AMD][AgentX] kimik3: L1=512GB — 1360GB breaks GPU IPC registration
hyukjlee Jul 30, 2026
463f1a9
[AMD][AgentX] kimik3: add c16 to the LMCache arm
hyukjlee Jul 30, 2026
f815885
[AMD][AgentX] kimik3: size LMCache L1 as a fraction of free /dev/shm …
hyukjlee Jul 31, 2026
aae1395
perf(amd): tune Kimi-K3 LMCache workers and eviction
hyukjlee Jul 31, 2026
95c37a6
test(amd): add Kimi-K3 LMCache c32
hyukjlee Jul 31, 2026
0aad1dd
[AMD][AgentX] kimik3: GMU 0.85 for c1 only, plus a c1 no-offload key
hyukjlee Jul 31, 2026
a4c104a
Merge branch 'main' into kimik3-agentic-smoke
seungrokj Aug 1, 2026
0b50bea
[AMD][AgentX] kimik3: reduce the PR to the DSpark arm only
functionstackx Aug 2, 2026
df1836a
refactor(amd): reduce this PR to the DSpark wrapper only
hyukjlee Aug 6, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 21 additions & 0 deletions benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,21 @@
#!/usr/bin/env bash
set -euo pipefail

# DSpark variant of kimik3_fp4_mi355x.sh. The MI355X launcher routes
# spec-decoding=mtp rows to this suffix, while the shared base recipe owns the
# model, AgentX replay, and eval plumbing.
#
# Keep this wrapper aligned with the upstream AMD Kimi-K3 DSpark reproducer.
# The first AgentX validation is deliberately GPU-only at c1 so a server or
# kernel failure has nowhere else to hide.
export SPEC_DECODE=true
export KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-auto}"
export GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.95}"
export MAX_NUM_SEQS="${MAX_NUM_SEQS:-16}"
export EVAL_MAX_NUM_SEQS="${EVAL_MAX_NUM_SEQS:-128}"
export MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-4096}"
export LANGUAGE_MODEL_ONLY="${LANGUAGE_MODEL_ONLY:-false}"
export PREFIX_CACHING="${PREFIX_CACHING:-auto}"
export ENFORCE_EAGER="${ENFORCE_EAGER:-false}"

exec "$(dirname "$0")/kimik3_fp4_mi355x.sh" "$@"