Repository navigation
Measure matched four-model native, PD and data-parallel HTTP serving - #251
Draft
gspark-etri wants to merge 4 commits into
Draft
gspark-etri wants to merge 4 commits into
gspark-etri wants to merge 4 commits into
Conversation
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Background
Existing lifecycle checks did not establish PD speedup under matched cache/chunk policies or a same-two-GPU data-parallel comparison.
Changes
Add an HTTP benchmark with isolated production server processes, native/PD/DP2 modes, separate warmup and token checks, raw SSE request timing, whole-batch throughput and output parity. Include commands, limitations and forward/reverse four-model blocks' raw evidence.
Verification
Accounting test1passed; Ruff and diff/sensitive checks clean. Hardware: two A10080GB PCIe, driver580.178.04; core d0f6f12 plus harness and HTTP candidate9e442cd. Four models (Qwen3-0.6B/4B, KONI-8B, Gemma3-1B), inputs128/8192, output32, concurrency1/4, three measured repetitions:24model/block/mode configurations,288batches, all parity passed. JSON records software versions and harness hash.
At8192input/c4, native/PD/DP2 tok/s respectively: Qwen0.6B20.43/22.37/40.15; Qwen4B12.86/13.25/25.17; KONI12.38/14.47/24.69; Gemma20.23/23.11/39.66.
Compatibility and limits
Experiment tooling only. First block is ordered; a short CPU regression run overlapped KONI and is documented. Quiet reverse-order replication also passes but PD versus native varies (Gemma23.11->17.79,Qwen0.6B22.37->18.41,Qwen4B13.25->15.12,KONI14.47->14.02tok/s at8192/c4). DP2 remains ahead for all four. The output128 extension completed 216 additional batches, all parity passed. At input8192/c8, native/PD/DP2 tok/s: Qwen0.6 65.74/74.66/120.65; Qwen4 44.32/46.78/80.42; KONI 42.87/50.13/80.45; Gemma 64.08/73.72/112.83. Heterogeneous results are tracked separately in #252; next-request prefetch in #255. DP2 has two decoder replicas/eight aggregate slots versus PD's four; both use two GPUs. First SSE event latency and estimated TPOT are not exact per-token timings. No within-request compute/copy overlap is implemented by this Torch baseline. No model-quality or final speedup claim. This PR does not establish heterogeneous or native-vendor-engine performance.
Related Issue
Refs #248; not closing the broader measurement task. Stacked on #247. Stage metric export is separate #250.