[WIP][experimental] multi turn chat benchmark#821
Draft
Conversation
* fix AITER flags for v0.14.0 release * drop mi325 triton gemm env var * Add changes to perf changelog
…wont be erroneous negative diff [skip-sweep] (#571)
* remove assign * initial * update perf * fix perf changelog * trigger test sweep * trigger test sweep pt 2 * rebase for evals only * Update perf-changelog.yaml * remove newline * update perf changelog --------- Co-authored-by: Cam Quilici <cjquilici@gmail.com>
* b300 srt slurm * update generated srtslurm yaml Signed-off-by: jthomson04 <jothomson@nvidia.com> * fix image * add uv and sqsh file * change partition * change slurm account * use regular srt Signed-off-by: jthomson04 <jothomson@nvidia.com> * update perf changelog Signed-off-by: jthomson04 <jothomson@nvidia.com> * fix runner Signed-off-by: jthomson04 <jothomson@nvidia.com> * correct account Signed-off-by: jthomson04 <jothomson@nvidia.com> * qos support Signed-off-by: jthomson04 <jothomson@nvidia.com> * fix get checkout Signed-off-by: jthomson04 <jothomson@nvidia.com> * update runner label and partition * undo branch checkout Signed-off-by: jthomson04 <jothomson@nvidia.com> * debug info Signed-off-by: jthomson04 <jothomson@nvidia.com> * cleanup logging Signed-off-by: jthomson04 <jothomson@nvidia.com> * use local model dir Signed-off-by: jthomson04 <jothomson@nvidia.com> * checkout specific commit Signed-off-by: jthomson04 <jothomson@nvidia.com> --------- Signed-off-by: jthomson04 <jothomson@nvidia.com> Co-authored-by: Sahithi Chigurupati <schigurupati@nvidia.com> Co-authored-by: Sahithi Chigurupati <chigurupati.sahithi@gmail.com>
…wont be erroneous negative diff [skip-sweep] (#577)
* Update SGLang Docker Image for MI355 to v0.5.8 1. activate FP8 KV cache 2. use the MLA persistent kernel * Do not activate FP8 KV cache and the MLA persistent kernel explicitly * Add config-keys (v0.5.5.post3 --> v0.5.8) * Update perf-changelog.yaml with key fix description for v0.5.8 Add description: Disables mla persistent kernel when not using fp8 kv_cache Co-authored-by: functionstackx <functionstackx@users.noreply.github.com> --------- Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com> Co-authored-by: functionstackx <functionstackx@users.noreply.github.com> Co-authored-by: functionstackx <47992694+functionstackx@users.noreply.github.com>
This reverts commit dea7f48.
30s default to 300s
* chore: save server long as artifact after single node runs * test flaky eval * test flaky eval * test flaky eval * rebase * rebase pt 2 * add trap to upload server logs on exit * rebase pt 3 * make server log in gha workspace * export result filename at runtime so it is present * revert perf changelog
* chore: add pre-merge check for newline in perf-changelog.yaml Add a validation step in run-sweep.yml that ensures perf-changelog.yaml ends with a newline character. This prevents negative diff issues in subsequent PRs when the file is appended to. Closes #578 Co-authored-by: Cameron Quilici <cquil11@users.noreply.github.com> * test * change logic of newline check * trigger test check * remove test perf changelog --------- Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com> Co-authored-by: Cameron Quilici <cquil11@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When DURATION is set, generates a 10k conversation pool and runs for that duration with grace-period=0. When unset, runs to completion. Added --dataset-sampler shuffle to avoid sequential ordering bias. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Runs analyze_benchmark_distributions.py after AIPerf completes to generate turn count, ISL/OSL distribution stats and plots. Results uploaded as artifacts for verification against Qwen trace profile. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Allows running sweeps with different model precisions by passing -f precision='fp8' to select the correct benchmark script. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…g, fix CUDA arch - Remove compilation-config (Blackwell-specific custom ops) - TORCH_CUDA_ARCH_LIST=9.0 (Hopper, not 10.0 Blackwell) - Remove --attention-config.use_trtllm_attention=0 (Blackwell-specific) - Add --disable-log-requests Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Synthetic workload: 10 turns (stddev 1), 2000 ISL (stddev 200), 500 OSL (stddev 50), 2s think time (stddev 500ms). No dataset file — uses AIPerf built-in synthetic generation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
vLLM docs: "When TP > 1, this is the total buffer size summed across all TP ranks." We were dividing by TP, giving each rank less offload space than intended. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The vLLM docs say "when TP > 1, this is the total buffer size summed across all TP ranks" — meaning the value is per-rank and the total is the sum. Original calculation was correct. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
vLLM passes kv_offloading_size * (1 << 30) directly as cpu_bytes_to_use. The per-rank division happens internally via world_size in the block cost calculation. Our division by TP was double-dividing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Error/timeout records may not have input_sequence_length or output_sequence_length in their metrics, causing KeyError. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The min(6000, ...) was hiding all requests above 6K tokens, making multi-turn context growth invisible in the plot. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New row: max ISL per conversation (final context size), total OSL per conversation, and max ISL vs turn count scatter. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
conversation_id is the template ID which gets reused across sessions. x_correlation_id is unique per session. Without this fix, sessions sharing the same template appeared as one conversation with duplicate turn indices, inflating turn counts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Same workload as H200 homogeneous (10 turns, 2k ISL, 500 OSL) but with Blackwell-specific config: CUDA 10.0, compilation config, trtllm attention disabled. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Enable --steady-state-prefill in homogeneous benchmark scripts. Update aiperf submodule with steady-state prefill implementation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two new features for benchmark_serving_multi_turn.py: 1. --synthetic mode: generates multi-turn conversations using Shakespeare corpus with configurable ISL/OSL/turns distributions (mean + stddev). Self-contained, no AIPerf dependency. 2. --steady-state-prefill: actually runs prefill turns through the server before benchmarking starts. Each client's first conversation is assigned a staggered starting turn and turns 0..N-1 are executed to warm the KV cache. Unlike --virtual-history which just skips turns (cold KV), this produces warm cache at benchmark start. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Same workload as AIPerf homogeneous (10 turns, 2k ISL, 500 OSL) but uses benchmark_serving_multi_turn.py with --synthetic and --steady-state-prefill. One conversation per client (max-active=1). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Custom client doesn't have --ignore-eos, --save-result, --result-dir, --result-filename. Removed these flags. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sets --limit-min-tokens 500 --limit-max-tokens 500 to ensure the model generates exactly 500 tokens per turn (equivalent to ignore_eos in AIPerf). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Setting --limit-min-tokens 0 --limit-max-tokens 0 tells the client to read the assistant message token count from the dataset and use it as both min_tokens and max_tokens per request. This gives per-request ignore_eos behavior matching the synthetic OSL distribution. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Passes ignore_eos=true in the API request payload to force generating until max_tokens. Works the same way as AIPerf's --extra-inputs ignore_eos:true. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
max_active_conversations is divided by num_clients to get per-client limit. Setting it to 1 with 512 clients gives 0, which fails. Set to USERS so each client manages 1 conversation at a time. Co-Authored-By: Claude Opus 4.6 (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
5 participants
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.
No description provided.