[NV] Add B300 DeepSeek-V4-Pro AgentX with SGLang DSpark6 / 使用 SGLang DSpark6 新增 B300 DeepSeek-V4-Pro AgentX - #2624

Closed

weireweire wants to merge 22 commits into

mainfrom

perf/dsv4-b300-dspark-agentx

Commits on Aug 16, 2026

perf(agentx): refresh GB300 DSV4 with DSpark6
weireweire
committed
83e2635
fix(agentx): keep DSpark synthetic AL out of eval
weireweire
committed
8a4a0a0
perf(b300): add DSpark6 Dynamo AgentX recipes
weireweire
committed
927f3d3
docs(perf): register B300 DSpark6 AgentX sweep
weireweire
committed
e9c00d8
fix(b300): make DSpark6 AgentX PR independent
weireweire
committed
e28fa42
perf(dsv4): calibrate DSpark6 acceptance to 3.77
weireweire
committed
1cbf9c3
fix(ci): place B300 AgentX entry at YAML top level
weireweire
committed
4b038b9
fix(ci): fit B300 DEP8 HiCache within host DRAM
weireweire
committed
4324132
ci: retrigger B300 AgentX sweep after HiCache fix
weireweire
committed
2be8f91
Merge remote-tracking branch 'origin/main' into perf/dsv4-b300-dspark-agentx
weireweire
committed
d39f736

Commits on Aug 17, 2026

fix(agentx): emit one staged tokenizer argument
weireweire
committed
2366223
Preserve B300 AgentX affinity across long sweeps
weireweire
committed
831b912
Match B300 DP-local prefill chunks to GB300
weireweire
committed
abf9155
perf: leave prefill activation headroom on B300 DEP8
weireweire
committed
8d66f6c
fix: use staged B300 DSpark checkpoint
weireweire
committed
3cf6a42
cleanup: use native B300 DSpark model metadata
weireweire
committed
b19ddad
fix: retain mainline trust-remote-code setting
weireweire
committed
b62a50c
fix(agentx): pin B300 reasoning effort high
weireweire
committed
8042c25
fix(agentx): align B300 tokenizer with 0813 model
weireweire
committed
8bdb32e
Merge remote-tracking branch 'origin/main' into perf/dsv4-b300-dspark-agentx
weireweire
committed
1d3656e
Enable B300 DP prefill phase alignment
weireweire
committed
78c1bd6
fix(agentx): handle B300 overlength traces
weireweire
committed
9da2042