Repository navigation
[DeepSeek] V4.1-Flash: TP2 with Engram offload on B200 - #994
Closed
functionstackx wants to merge 2 commits into
Closed
functionstackx wants to merge 2 commits into
functionstackx wants to merge 2 commits into
Conversation
Default the single-node TP strategy to 2 GPUs on B200 as well as B300 and move the Engram tables to pinned host DRAM on both. B200 (180 GB per GPU) also caps --max-num-batched-tokens at 4096 and --max-cudagraph-capture-size at 512 with --max-num-seqs 256, the settings under which the InferenceX AgentX sweep completed concurrency 1-128 (SemiAnalysisAI/InferenceX#3216, run 35460277352). The guide's 4xB300 section becomes a B200/B300 TP2 section. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Contributor
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request updates the configuration for the DeepSeek-V4.1-Flash model by adding support for B200 hardware, including specific memory and performance optimizations such as CPU offloading for Engram tables and batched token caps. Additionally, the documentation has been updated to reflect these changes. I have no feedback to provide.
…300 one Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
functionstackx
changed the title
Contributor
Author
This branch was successfully deployed
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Upstream has since shipped the B300 TP2 + Engram CPU offload override, so this PR now adds only the B200 counterpart:
strategy_overrides.single_node_tp.tp.b200: 2beside the existingb300: 2.b200override (takes precedence over the sharedblackwelldefaults):--engram-config '{"cpu_offload":true}',--max-num-batched-tokens 4096,--max-cudagraph-capture-size 512,--max-num-seqs 256.Why these levers: TP2 on 180 GB B200 with the Engram tables offloaded still leaves ~145 GiB of resident weights per GPU. At the upstream 16384 batched tokens the sparse-attention indexer's
[batched-tokens, 1M]fp8 logits buffer is 32 GiB and graph capture above 512 tokens pushed the KV budget negative. With batched tokens capped at 4096 and capture stopped at 512, the InferenceX AgentX sweep completed concurrency 1 through 128 with evals green (SemiAnalysisAI/InferenceX#3216, run https://github.com/SemiAnalysisAI/InferenceX/actions/runs/35480787586). FlashInfer's autotune picked an invalid MXFP8 tactic below 16 sequences, hence the--max-num-seqsfloor noted in the comment.Validation
node scripts/build-recipes-api.mjsrenders the B200 page with TP2, the Engram offload flag and the caps.mainafter [DeepSeek] Update V4 indexer cache dtype configuration #992 and the B300 TP2 change; the B300 pieces of the original PR are dropped in favour of upstream's.🤖 Generated with Claude Code