workload: add guide_p2p-kv-cache-sharing_1 profile (document Q&A)#1656
Draft
nilig wants to merge 4 commits into
Draft
workload: add guide_p2p-kv-cache-sharing_1 profile (document Q&A)#1656nilig wants to merge 4 commits into
nilig wants to merge 4 commits into
Conversation
Dedicated inference-perf profile for the p2p-kv-cache-sharing guide's headline benchmark: 192 conversations with 48K-token private document prefixes, 6 short Q&A turns each, 128 concurrent. Sized so the corpus oversubscribes a 14-16 pod H200 fleet's aggregate GPU KV; scale num_conversations to the target fleet. Signed-off-by: nilig <nili.ifergan@gmail.com>
This was referenced Jul 17, 2026
The --spec guides/p2p-kv-cache-sharing entry point and the scenario it references complete the per-guide file set alongside the workload profile. The scenario deploys via the upstream guide's own manifests (kustomize mode), so the stack definition is not mirrored here. Signed-off-by: nilig <nili.ifergan@gmail.com>
A fresh capacity measurement on the guide's pinned config put per-pod GPU KV at ~1.22M tokens, not the ~0.48M this comment assumed - at the corrected figure the 9.2M-token corpus fits inside the fleet's aggregate GPU cache, so what drives placement here is per-pod queueing under concurrency, not corpus size versus aggregate capacity. See llm-d#2067 for the corresponding guide/report fix. Signed-off-by: nilig <nili.ifergan@gmail.com>
…load profile Matches the guide benchmark report's uniform-pool parameters: 128 shared 48K-token prefixes, 256-token questions, 64-token outputs, constant-rate stages 4-24 req/s. Completes the guide's dedicated profile set alongside the docQA headline profile. Signed-off-by: nilig <nili.ifergan@gmail.com>
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.
Part of llm-d/llm-d#2066 (companion to llm-d/llm-d#2067).
Dedicated inference-perf profile for the p2p-kv-cache-sharing guide's headline benchmark: 192 conversations with 48K-token private document prefixes, 6 short Q&A turns each, 128 concurrent. Sized so the corpus oversubscribes a 14-16 pod H200 fleet's aggregate GPU KV; the profile documents scaling
num_conversationsto other fleets. The guide's benchmarking README references this profile in its run instructions.