Repository navigation
[K3.1-Pro] Day-0 support SGLang on B200 #4801
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Merged
Merged
Changes from all commits
Commits
Show all changes
6 commits
Select commit
Hold shift + click to select a range
24d5293
feat(kimik3.1): add Kimi K3.1 Pro B200 SGLang AgentX config
cquil11 e80fb62
chore(kimik3.1): link the B200 SGLang changelog entry to its PR
cquil11 e349388
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-b200-sglang
cquil11 256ba5d
fix(kimik3.1): turn off Dynamo conversation routing for the B200 SGLa…
cquil11 5bb0810
Merge remote-tracking branch 'origin/main' into cam/kimik3.1-b200-sglang
cquil11 4ef918d
chore: refresh PR #4801 for sweep reuse [skip-sweep]
cquil11 File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
168 changes: 168 additions & 0 deletions
168
...benchmarks/multi_node/srt-slurm-recipes/kimik3.1/sglang/b200-fp4/agentx/agg-variants.yaml
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,168 @@ | ||
| schema: 2 | ||
|
|
||
| base: | ||
| identity: | ||
| model: {} | ||
| container: {} | ||
| resources: | ||
| gpu_type: b200 | ||
| gpus_per_node: 8 | ||
| slurm: | ||
| time_limit: '8:00:00' | ||
| dynamo: | ||
| install: false | ||
| # One aggregate worker serves the public port itself; no router is needed. | ||
| frontend: | ||
| type: sglang | ||
| enable_multiple_frontends: false | ||
| observability: | ||
| enabled: false | ||
| tachometer: | ||
| enabled: false | ||
| engine: sglang | ||
| roles: | ||
| agg: | ||
| # One 8xB200 node cannot hold the ~1.56 TB MXFP4 checkpoint. | ||
| nodes: 2 | ||
| workers: 1 | ||
| gpus: 16 | ||
| env: | ||
| TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800' | ||
| PYTHONUNBUFFERED: '1' | ||
| PYTHONNOUSERSITE: '1' | ||
| # Outlast AIPerf's pooled connections past Uvicorn's keep-alive. | ||
| SGLANG_TIMEOUT_KEEP_ALIVE: '1800' | ||
| args: | ||
| served-model-name: moonshotai/Kimi-K3.1-Pro | ||
| trust-remote-code: true | ||
| # The SGLang Kimi K3.1 Pro cookbook's two-node B200 layout. | ||
| tp-size: 16 | ||
| dcp-size: 16 | ||
| mem-fraction-static: 0.85 | ||
| weight-loader-prefetch-checkpoints: true | ||
| model-loader-extra-config: '{"enable_multithread_load": true}' | ||
| kv-cache-dtype: fp8_e4m3 | ||
| # The K3.1 week-1 precision policy requires FP32 KDA recurrent state. | ||
| mamba-ssm-dtype: float32 | ||
| reasoning-parser: kimi_k3 | ||
| tool-call-parser: kimi_k3 | ||
| # DFLASH verifies the current token plus 7 drafts per step. | ||
| speculative-algorithm: DFLASH | ||
| speculative-draft-model-path: RadixArk/Kimi-K3.1-Dflash2-Preview | ||
| speculative-num-draft-tokens: 8 | ||
| speculative-use-block-verification: true | ||
| speculative-draft-kv-cache-dtype: fp8_e4m3 | ||
| speculative-draft-attention-backend: cutedsl_mla | ||
| # Without ReplaySSM DFLASH keeps one KDA state per draft token and the KV pool collapses. | ||
| enable-linear-replayssm-spec: true | ||
| watchdog-timeout: 3600 | ||
| enable-metrics: true | ||
| enable-cache-report: true | ||
| health_check: | ||
| max_attempts: 1440 | ||
| interval_seconds: 10 | ||
| benchmark: | ||
| env: | ||
| AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:' | ||
| # Warmup can leave bytes unacknowledged past AIPerf's 30 s default. | ||
| AIPERF_HTTP_TCP_USER_TIMEOUT: '900000' | ||
| AGENTIC_WARMUP_GRACE_PERIOD: '3600' | ||
| # SGLang serves the port itself; conversation-aware routing needs a Dynamo frontend. | ||
| AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0' | ||
| sbatch_directives: | ||
| mem: '0' | ||
| srun_options: | ||
| mem: '0' | ||
| container-remap-root: '' | ||
|
|
||
| # Admission is 2x CONC. Seven drafts through CONC 14; four (the draft's native block of 5) | ||
| # from CONC 24, where the verify batch grows with the running requests. | ||
| override_tp16_c1: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 2 | ||
| cuda-graph-max-bs-decode: 2 | ||
|
|
||
| override_tp16_c4: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 8 | ||
| cuda-graph-max-bs-decode: 8 | ||
|
|
||
| override_tp16_c8: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 16 | ||
| cuda-graph-max-bs-decode: 16 | ||
|
|
||
| override_tp16_c14: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 28 | ||
| cuda-graph-max-bs-decode: 28 | ||
|
|
||
| override_tp16_c24: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 48 | ||
| cuda-graph-max-bs-decode: 48 | ||
| speculative-num-draft-tokens: 5 | ||
|
|
||
| override_tp16_c32: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 64 | ||
| cuda-graph-max-bs-decode: 64 | ||
| speculative-num-draft-tokens: 5 | ||
|
|
||
| override_tp16_c48: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 96 | ||
| cuda-graph-max-bs-decode: 96 | ||
| speculative-num-draft-tokens: 5 | ||
|
|
||
| override_tp16_c56: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 112 | ||
| cuda-graph-max-bs-decode: 112 | ||
| speculative-num-draft-tokens: 5 | ||
|
|
||
| override_tp16_c72: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 144 | ||
| cuda-graph-max-bs-decode: 144 | ||
| speculative-num-draft-tokens: 5 | ||
|
|
||
| override_tp16_c80: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 160 | ||
| cuda-graph-max-bs-decode: 160 | ||
| speculative-num-draft-tokens: 5 | ||
|
|
||
| override_tp16_c96_nospec: | ||
| roles: | ||
| agg: | ||
| args: | ||
| max-running-requests: 192 | ||
| cuda-graph-max-bs-decode: 192 | ||
| speculative-algorithm: null | ||
| speculative-draft-model-path: null | ||
| speculative-num-draft-tokens: null | ||
| speculative-use-block-verification: null | ||
| speculative-draft-kv-cache-dtype: null | ||
| speculative-draft-attention-backend: null | ||
| enable-linear-replayssm-spec: null |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
🔴 If this merges before #4798, all ten DFLASH rows fail at binding with
No golden curve for model prefix 'kimik3.1' and method 'dflash'. Only the c96 no-spec row would launch. identities.yaml has no kimik3.1 dflash identity, and this diff does not add one. Fix: add thekimi-k3.1-pro__dflash__kimi-k3.1-dflash2-preview__default-blockidentity and its curve in this change. Otherwise, gate the merge on #4798 or leave the DFLASH rows out until it lands.Why this was flagged
The condition is merging this PR before the stacked #4798. Every search-space row with
spec-decoding: draft_model(nvidia-master.yaml:4817-4944) resolves a golden AL curve by model prefixkimik3.1and methoddflash. infx/golden_al_distribution/identities.yaml has kimik3.1 only under method dspark (line 105-106). The only dflash identity there is for musespark (line 170-171). identity.py:260 therefore raises ValueError("No golden curve for model prefix 'kimik3.1' and method 'dflash'") for 10 of the 11 rows, so the sweep produces only the c96 no-spec point. The PR description acknowledges the dependency, but nothing in this diff enforces the merge order.Verification: The diff does not touch infx/golden_al_distribution/identities.yaml. agg-variants.yaml sets
speculative-algorithm: DFLASH(line 50). identities.yaml selects kimik3.1 only withmethod: dspark(line 106), so identity.py:260 raisesValueError. Result: all tenspec-decoding: draft_modelrows fail to launch.