feat(evaluation): add topology grouping + technique-category filter - #9
Merged
Merged
Conversation
Two fixes after the first real recipe-comparison ingest of gemma4 + the
two new sweeps (1×H100 mnbt sweep, 2×H100 spec-decoding):
1. Topology was collapsing. groupKey only included
(model, hw, framework, precision, isl, osl, conc) — so 1× H100
(TP=1) and 2× H100 (TP=2) rows merged into a single "h100" row group
and the user couldn't tell them apart. Now the groupKey also
includes num_prefill_gpu, num_decode_gpu, prefill_tp, decode_tp,
disagg. New "Topology" column renders "1× (TP=1)" / "2× (TP=2)"
for non-disagg and "{P}P+{D}D (TP=p/d)" for disagg layouts.
2. Spec-decoding and batch-size tuning runs were mixing in the same
view. Added a filter chip group at the top of the page driven by
`categorizeTechniques`:
- baseline (techniques = {})
- spec-decoding (spec_method set)
- batch-size (max_num_batched_tokens set)
- kv-cache (kv_cache_dtype set)
- prefix-cache (prefix_cache set)
- other (anything else)
Only chips with at least one row are rendered. When a non-"all"
chip is active, baseline rows still show through so the user keeps
a reference variant.
Tests: 10/10 pass (added 2 new — topology-split + categorize).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
After ingesting the two new gemma4 runs (1×H100 mnbt sweep + 2×H100 spec-decoding), two issues showed up in the recipe table:
1. Topology was collapsing
The groupKey was
(model | hw | framework | precision | isl | osl | conc). 1× H100 (TP=1) and 2× H100 (TP=2) rows both produced `groupKey = "gemma4|h100|vllm|fp8|..."` and merged into one "h100" row group with no visual distinction. Fix:2. Different recipe axes were mixing in one view
The 1×H100 run varies `max_num_batched_tokens`. The 2×H100 run varies `num_speculative_tokens`. Side-by-side in the same table = noise. Added a chip group:
```
Filter by knob: [All] [Spec decoding] [Batch size] [KV cache] [Prefix cache] [Baseline only]
```
Categories are derived from `techniques` via `categorizeTechniques()`. Only chips with data are rendered. Baseline rows show through any active non-"all" filter so the user always has a reference variant.
Test plan
🤖 Generated with Claude Code