feat: add --remote flag for cloud/remote API benchmarking#12
Merged
Conversation
added 3 commits
May 22, 2026 22:51
- Add --remote CLI flag (auto-detected from non-localhost endpoints) - Reweight overall scoring for remote runs: 65% quality + 35% reliability (drops 20% speed weight since local tok/s is meaningless for cloud APIs) - Skip model validation when /v1/models returns empty (common for DashScope, etc.) - Show 'N/A (cloud)' for speed in summary output - Add is_remote field to BenchmarkRun for leaderboard filtering - Display remote mode indicator in console report
- Add chat_streaming() to openai_compat provider — measures TTFT and effective tok/s via SSE streaming (what cloud users actually feel) - Add cloud speed scoring curve: 60% TTFT + 40% effective tok/s (TTFT 200ms->100, 500ms->86, 2000ms->40; tok/s 40->60, 100->85) - Use streaming for speed tasks when remote=true - Include cloud speed in overall when available (0.50Q + 0.25S + 0.25R) - Show TTFT + tok/s in summary output for remote runs - Add is_cloud_speed metadata to speed task results
- Track first reasoning token as TTFT (what users see first) - Track first content token separately for throughput calculation - Subtract reasoning_tokens from completion_tokens for content tok/s - Measure content throughput from first content token, not TTFT - Fixes qwen3.7-max and similar reasoning models showing inflated token counts and near-zero tok/s
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Adds first-class support for benchmarking cloud/remote LLM APIs (DashScope, OpenRouter, Together, OpenAI, etc.) without the local tok/s speed score tanking the overall.
Problem
The current scoring formula is
0.55 × quality + 0.20 × speed + 0.25 × reliability. For cloud APIs,speed_scoreis always 0 because the speed suite measures local inference throughput (tok/s). This drops cloud models to ~70 overall even with 90+ quality scores.Changes
--remoteCLI flag — explicit opt-in, also auto-detected when endpoint is not localhost0.65 × quality + 0.35 × reliability(speed weight redistributed)is_remotefield onBenchmarkRun— enables leaderboard to filter/separate cloud vs local runs/v1/modelsreturns empty (DashScope and some other providers don't implement this endpoint)☁ remote/cloudindicator andN/A (cloud)for speedExample
Before (without
--remote):After (with
--remote):Files changed
bench_loop/cli.py— added--remoteflagbench_loop/models.py— addedis_remotefield + reweightedcompute_aggregates()bench_loop/runner/orchestrator.py— auto-detect remote, skip empty model list validationbench_loop/report/console.py— remote indicator + N/A speed display