[KimiK3][AgentX]: H200 KimiK3 Day 0 support - #2353

Merged

cquil11 merged 92 commits into

mainfrom

agent/kimik3-h200-agentx

Aug 7, 2026

Commits on Jul 29, 2026

feat(agentx): add Kimi K3 H200 day-0 support
cquil11
committed
9b1e9b2
chore: link Kimi K3 H200 changelog to PR
cquil11
committed
16d44ef
refactor(runners): add pluggable Slurm multi-node connector
cquil11
committed
18a61bd
refactor(runners): simplify native multi-node launch
cquil11
committed
c31b634
fix(kimik3): use default model context length
cquil11
committed
f4279ac
docs(kimik3): trim benchmark changelog
cquil11
committed
902951a
refactor(kimik3): trust benchmark environment
cquil11
committed
4be37db
fix(h200): set native vLLM port
cquil11
committed
b12f1ae
fix(kimik3): unblock H200 model loading
cquil11
committed
2d8d435
fix(kimik3): align rendezvous host with Slurm rank zero
cquil11
committed
ec0c20d
fix(kimik3): provision full-context H200 topology
cquil11
committed
bbbbc3d
fix(kimik3): stream H200 checkpoint loading
cquil11
committed
428e72d
fix(kimik3): fit full context with pipeline parallelism
cquil11
committed
2325df7
fix(kimik3): use supported model runner
cquil11
committed
e29e380
fix(h200): preserve server logs on cancellation
cquil11
committed
24fdf3f
fix(kimik3): preserve full context with DCP
cquil11
committed
9baa292
fix(kimik3): fit H200 startup memory budget
cquil11
committed
00acd46
chore(h200): name the vLLM server log clearly
cquil11
committed
8e5f309
test(kimik3): reproduce pipeline serving path
cquil11
committed
8ddd1b9
fix(kimik3): use two-node reference topologies
cquil11
committed
a73f110
fix(h200): reuse host git for agentic clients
cquil11
committed
8cd634a
test(kimik3): probe two-node DCP serving
cquil11
committed
e1f7d83
test(kimik3): probe two-node TEP serving
cquil11
committed
0335cd9
fix(kimik3): stream TEP and DEP weight loading
cquil11
committed
08d7400
fix(h200): expose all local DP endpoints
cquil11
committed
ccb01fa
test(kimik3): probe full-context TP16 serving
cquil11
committed
9609bca
fix(kimik3): leave room for distributed startup
cquil11
committed
e5cdcbb
fix(kimik3): stream all multi-node weight loads
cquil11
committed
0fda2ca
fix(h200): stop failed engines during readiness
cquil11
committed
b92ce48
test(kimik3): probe four-node DP16 serving
cquil11
committed
2b8c47f
test(kimik3): run PP with the V1 model runner
cquil11
committed
18e2b3f
fix(kimik3): allow distributed worker overhead
cquil11
committed
b7e58f3
perf(kimik3): use fast loading for PP
cquil11
committed
b3b628d
fix(kimik3): use bf16 kv cache for PP
cquil11
committed
f2d1243
fix(kimik3): reserve PP runtime memory
cquil11
committed
83c9ca5
perf(kimik3): add PP to EP strategies
cquil11
committed
5fc99b0
fix(kimik3): reserve topology workspace
cquil11
committed
a65e0b9
fix(kimik3): reduce latency prefill chunk
cquil11
committed
f2090e7
fix(kimik3): align H200 overrides with official recipe
cquil11
committed
adc1109
fix(kimik3): disable SM100-only tail fusion on H200
cquil11
committed
d33f844
fix(kimik3): use one shared API port per DP node
cquil11
committed
a4c259b
fix(kimik3): leave throughput prefill headroom
cquil11
committed
deae74e
perf(kimik3): restore Day 0 sweep ranges
cquil11
committed
d0eefab
fix(kimik3): preserve throughput activation headroom
cquil11
committed
6e98eed
fix(h200): guard cleanup before server startup
cquil11
committed
7c2bcd4
fix(h200): snapshot native server logs during runs
cquil11
committed
f99af81
fix(kimik3): reduce throughput prefill workspace
cquil11
committed
724ad15
fix(kimik3): preserve throughput collective headroom
cquil11
committed
8abfa19
fix(kimik3): allow long throughput warmup drains
cquil11
committed
2c2d6de
fix(kimik3): cap H200 throughput concurrency
cquil11
committed
572b85b
fix(kimik3): use KV-safe throughput concurrency
cquil11
committed
caa2e20
fix(kimik3): add throughput tensor parallel headroom
cquil11
committed
3e27396
refactor(kimik3): use srt-slurm direct vLLM
cquil11
committed
2d0629d
fix(kimik3): disable SM100 tail fusion on H200
cquil11
committed
8ee6ab3
perf(kimik3): add high-concurrency KV offload
cquil11
committed
7ccb00e
perf(kimik3): densify H200 concurrency sweep
cquil11
committed
ec22d4b
Update perf-changelog.yaml
cquil11
committed
4c9dcdf
perf(kimik3): enable DSpark on H200
cquil11
committed
1737359
fix(kimik3): replace PP with pure H200 topologies
cquil11
committed
5c03d8f
Merge origin/main into H200 Kimi K3 branch
cquil11
committed
735784f
perf(kimik3): use fit-safe H200 hybrid topologies
cquil11
committed
631fa8a
perf(kimik3): remove duplicate H200 throughput arm
cquil11
committed
d451287
perf(kimik3): enable Model Runner V2 on H200
cquil11
committed
9c8643c

Commits on Jul 30, 2026

chore: merge origin/main into Kimi K3 H200 support
cquil11
committed
d2827ca
fix(kimik3): use FlashMLA for H200 DSpark
cquil11
committed
174fa1f
Merge main into H200 Kimi-K3 AgentX branch
cquil11
committed
d4d6a47
fix(kimik3): auto-select H200 DSpark draft attention
cquil11
committed
bd4b520
perf(kimik3): align H200 runtime limits
cquil11
committed
87b21b2
fix(kimik3): preserve native H200 capacity
cquil11
committed
3ac5e57
fix(kimik3): bypass unstable Hopper router batches
cquil11
committed
03cb89e
perf(kimik3): report prompt cache reads
cquil11
committed
94f6d36
Merge main into H200 Kimi-K3 AgentX branch
cquil11
committed
0aac694
fix(agentx): align H200 AIPerf replay
cquil11
committed
40d6f02
fix(agentx): let Dynamo report cached prompt tokens
cquil11
committed
ea11e1c

Commits on Jul 31, 2026

Merge main into H200 Kimi-K3 AgentX branch
cquil11
committed
6915490
fix(agentx): stabilize Kimi K3 DSpark on H200
cquil11
committed
fdc5a8f
fix(agentx): retain Kimi K3 RMS fusion guard
cquil11
committed
14d70db
fix(vllm): stabilize Kimi K3 TP8 DSpark
cquil11
committed
492a396
docs(config): clarify H200 DSpark levels
cquil11
committed
8cedb3d
fix(vllm): stabilize Kimi K3 TP16 DSpark
cquil11
committed
1e68b47
fix(agentx): pin Kimi K3 golden acceptance
cquil11
committed
4e60581
fix(vllm): serialize H200 Kimi K3 TP8 scheduling
cquil11
committed
e2554a5
Merge origin/main into H200 Kimi K3 AgentX branch
cquil11
committed
38c9093
perf(kimik3): restore direct vLLM cache metrics
cquil11
committed
0b3325c
fix(vllm): serialize Kimi K3 TP8 CUDA launches
cquil11
committed
0d9bbb0
fix(vllm): serialize Kimi K3 offload launches
cquil11
committed
e953279
fix(kimik3): serialize TP16 CUDA launches
cquil11
committed
432c601
chore(agentx): bump AIPerf harness
cquil11
committed
d58f2e9

Commits on Aug 2, 2026

chore(agentx): bump AIPerf barrier fixes
cquil11
committed
894719a

Commits on Aug 3, 2026

chore: merge origin/main into H200 Kimi K3 branch
cquil11
committed
ad5ee95
fix(agentx): remove duplicate trace idle cap flag
cquil11
committed
114c1bd

Commits on Aug 7, 2026

merge(main): resolve PR conflicts
cquil11
committed
d45431f