vllm-headroom 0.1.0__tar.gz

vllm_headroom-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

vllm_headroom-0.1.0/PKG-INFO

@@ -0,0 +1,195 @@
+Metadata-Version: 2.4
+Name: vllm-headroom
+Version: 0.1.0
+Summary: Estimate how much more load a vLLM deployment can take before it breaks its latency SLO
+License: Apache-2.0
+Project-URL: Homepage, https://github.com/haochen2049/vllm-headroom
+Project-URL: Source, https://github.com/haochen2049/vllm-headroom
+Project-URL: Issues, https://github.com/haochen2049/vllm-headroom/issues
+Project-URL: Example report, https://haochen2049.github.io/vllm-headroom/
+Keywords: vllm,llm,inference,capacity-planning,slo,queueing,simulation
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: System Administrators
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: System :: Benchmark
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Requires-Dist: httpx>=0.25
+Provides-Extra: dev
+Requires-Dist: pytest>=7; extra == "dev"
+Dynamic: license-file
+
+# vllm-headroom
+
+English | [中文](https://github.com/haochen2049/vllm-headroom/blob/main/README.zh-CN.md)
+
+**How much more traffic can your vLLM server take before it breaks its latency SLO?**
+
+`vllm-headroom` profiles a running vLLM server for about three minutes, fits a model of how long each
+engine step takes, and simulates continuous batching to predict p95/p99 latency at any load. It
+reports the highest request rate that still meets your SLO, with an uncertainty range, and checks its
+own predictions against load it measured but did not fit on.
+
+```text
+$ vllm-headroom analyze profile.json --slo ttft_p95=1 --slo tpot_p95=0.05 --current-rate 10 \
+      --whatif max_num_seqs=128 --whatif replicas=2
+saturation:             21.589 req/s
+max load meeting SLO:   20.827 req/s  (10-90%: 19.948-21.538; binding: ttft_p95)
+headroom vs current:    +108%  (10-90%: +99% to +115%)
+what-if max_num_seqs=128: max 27.745 req/s (saturation)
+what-if replicas=2: max 42.475 req/s (saturation)
+```
+
+It writes a self-contained HTML report with latency-versus-load curves, the backtest and the fitted
+model. Nothing leaves your machine. [See the full example report](https://haochen2049.github.io/vllm-headroom/) from a real 7B deployment.
+
+[![Example report: TTFT p95 versus load, with the SLO, the current load and the highest load that still meets the SLO](https://raw.githubusercontent.com/haochen2049/vllm-headroom/main/assets/report.png)](https://haochen2049.github.io/vllm-headroom/)
+
+> Status: alpha (0.1). Tested against vLLM 0.31 on one GPU type; see [Validation](#validation).
+> Not affiliated with the vLLM project.
+
+## Why
+
+Dashboards show what already happened, and forecasting tools extrapolate it. Neither tells you what
+happens after you change something, such as more traffic, a different `--max-num-seqs`, one fewer
+replica or a longer prompt mix, or how close you are to the point where latency falls off a cliff.
+Load testing every option is slow. A model fitted to *your* server answers these questions in seconds.
+
+## Quick start
+
+```bash
+pip install vllm-headroom
+
+# no GPU? profile a bundled simulated server and open the sample report (about 2-3 minutes)
+vllm-headroom demo -o demo_report.html
+```
+
+Against a real server, ideally an idle replica or a staging copy:
+
+```bash
+# 1. profile (about 3-4 minutes); pass the server's own settings if you know them
+vllm-headroom profile --url http://localhost:8000 \
+    --max-num-seqs 256 --max-num-batched-tokens 8192 -o profile.json
+
+# 2. analyse against your SLO and current traffic
+vllm-headroom analyze profile.json --slo ttft_p95=2 --slo tpot_p95=0.1 \
+    --current-rate 12 --workload my_requests.jsonl -o headroom.html
+
+# 3. compare configurations quickly
+vllm-headroom whatif profile.json --slo ttft_p95=2 \
+    --change max_num_seqs=512 --change replicas=3 --change kv_scale=0.5
+```
+
+`--workload` is a JSONL file with one request per line, `{"prompt_tokens": 812, "output_tokens": 230}`
+(OpenAI `usage` objects also work). Without it, `--prompt-mean/--output-mean` give a log-normal mix,
+and the default is the mix used during profiling. SLO metrics are `ttft` (time to first token),
+`tpot` (time per output token after the first) and `e2e` (end-to-end), each with any percentile:
+`ttft_p95=2`, `e2e_p99=30`, `tpot_p50=0.05`.
+
+## How it works
+
+**1. Profile.** Four phases against the OpenAI-compatible endpoint:
+
+| phase | what is sent | what it identifies |
+|---|---|---|
+| prefill sweep | single requests, prompts from 128 to 8192 tokens, idle server | prefill cost per token and fixed overhead |
+| decode waves | `c` identical requests at once, `c` = 1, 2, 4 ... 256, short and long prompts | step time versus batch size and context length |
+| admission probes | short requests sent while a wave decodes | how many steps a new request waits before it is scheduled |
+| validation | open-loop Poisson load at 30, 50 and 70% of predicted saturation | nothing: held out for the backtest |
+
+Prompts are random token ids, so prefix caching cannot shorten prefill and lengths are exact;
+`ignore_eos` fixes output lengths. Streaming chunk timestamps give inter-token gaps; only gaps with no
+prefill from another request are used for decode fitting. KV capacity is read from `/metrics`, and
+`max_num_seqs` is inferred from the waves if you do not pass it.
+
+**2. Fit.** Engine step time is modelled as
+
+```
+step = b0 + b_dec·(decoding sequences) + b_ctx·(their context tokens)
+          + b_pf·(prefill tokens) + b_pfatt·(prefill attention term)
+```
+
+fitted with robust least squares, plus a client-visible `overhead` and an `admit_lag` (vLLM V1 plans
+the next step while the current one runs, so a new request waits about one to two extra steps).
+A bootstrap over profiling requests gives parameter uncertainty.
+
+**3. Simulate.** A discrete-event simulator of vLLM-style continuous batching: chunked prefill under
+`max_num_batched_tokens`, `max_num_seqs`, KV capacity with recompute preemption, several replicas
+behind a least-loaded or round-robin router. Decode-only stretches are advanced in closed form, so a
+1,500-request run takes well under a second.
+
+**4. Report.** The maximum Poisson rate meeting every SLO (bisection), the saturation rate, latency
+curves with a 10–90% range across bootstrap refits, what-if comparisons and a backtest against the
+held-out validation load.
+
+## Validation
+
+One run so far: Qwen2.5-7B-Instruct on vLLM 0.31.0, one NVIDIA RTX PRO 6000 Blackwell (96 GB),
+`--max-num-seqs 64 --max-num-batched-tokens 4096 --gpu-memory-utilization 0.5`. The profile, report
+and check script are in [`examples/qwen2.5-7b-vllm0.31`](https://github.com/haochen2049/vllm-headroom/tree/main/examples/qwen2.5-7b-vllm0.31). The workload
+is the default log-normal mix (prompt mean 505, output mean 130 tokens). `max_num_seqs` was not passed
+and was inferred correctly (64).
+
+**Backtest** (held-out validation load from the same profiling run, never used for fitting):
+
+| load (of predicted saturation) | requests | TTFT p50 | TTFT p95 | E2E p50 | E2E p95 |
+|---|---|---|---|---|---|
+| 30% (6.5 req/s) | 161 | +8% | +7% | +2% | +2% |
+| 50% (10.8 req/s) | 268 | +9% | +12% | +2% | +2% |
+| 70% (15.1 req/s) | 376 | +8% | +14% | +3% | +4% |
+
+**Near saturation** (a separate later run against the same server with the fit frozen,
+`near_saturation_check.py`):
+
+| load | requests | TTFT p50 | TTFT p95 | E2E p50 | E2E p95 |
+|---|---|---|---|---|---|
+| 85% (18.5 req/s) | 833 | +7% | +5% | +3% | +5% |
+| 95% (20.7 req/s) | 931 | 0% | −19% | −3% | 0% |
+
+Entries are prediction error relative to measurement (positive = predicted slower). End-to-end
+latency is within 5% everywhere. TTFT is predicted somewhat pessimistically up to 85% load. At 95%
+load the TTFT tail is under-predicted by 19%, because close to saturation a small error in the rate
+moves the tail a lot. Treat the last few percent before saturation as unreliable and keep a margin.
+This is one model, one GPU and one workload; results on other setups may differ. If you run it on
+other hardware, please [share your backtest](https://github.com/haochen2049/vllm-headroom/issues/new?template=backtest.yml): it is the most useful contribution right now.
+
+The unit tests also check the simulator against a plain step-by-step reference implementation
+(identical results), recover known parameters from synthetic data, and run the full
+profile → fit → report pipeline over HTTP against the bundled simulated server.
+
+## Limits
+
+- One replica is profiled; multiple replicas are modelled as identical copies.
+- Arrivals are Poisson at a constant rate. Bursty traffic needs more headroom than reported.
+- Not modelled: prefix-cache hits, speculative decoding, LoRA adapter switching, multi-model serving,
+  disaggregated prefill/decode, pipeline-parallel bubbles, network between client and router.
+- Batch sizes and prompt lengths beyond what was profiled are extrapolated.
+- The uncertainty range reflects profiling noise, not every way production can differ.
+- Profiling sends real load. Run it on a replica that is not serving users.
+
+## Output files
+
+`profile` writes raw timings (no prompt or output text). `analyze` writes `headroom.html` and
+`headroom.json` with the fitted parameters, curves, headroom, backtest and what-if results.
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+pytest -q
+vllm-headroom mock-server --port 8001   # a simulated OpenAI-compatible server for experiments
+```
+
+## License
+
+Apache-2.0

vllm_headroom-0.1.0/README.md

@@ -0,0 +1,164 @@
+# vllm-headroom
+
+English | [中文](https://github.com/haochen2049/vllm-headroom/blob/main/README.zh-CN.md)
+
+**How much more traffic can your vLLM server take before it breaks its latency SLO?**
+
+`vllm-headroom` profiles a running vLLM server for about three minutes, fits a model of how long each
+engine step takes, and simulates continuous batching to predict p95/p99 latency at any load. It
+reports the highest request rate that still meets your SLO, with an uncertainty range, and checks its
+own predictions against load it measured but did not fit on.
+
+```text
+$ vllm-headroom analyze profile.json --slo ttft_p95=1 --slo tpot_p95=0.05 --current-rate 10 \
+      --whatif max_num_seqs=128 --whatif replicas=2
+saturation:             21.589 req/s
+max load meeting SLO:   20.827 req/s  (10-90%: 19.948-21.538; binding: ttft_p95)
+headroom vs current:    +108%  (10-90%: +99% to +115%)
+what-if max_num_seqs=128: max 27.745 req/s (saturation)
+what-if replicas=2: max 42.475 req/s (saturation)
+```
+
+It writes a self-contained HTML report with latency-versus-load curves, the backtest and the fitted
+model. Nothing leaves your machine. [See the full example report](https://haochen2049.github.io/vllm-headroom/) from a real 7B deployment.
+
+[![Example report: TTFT p95 versus load, with the SLO, the current load and the highest load that still meets the SLO](https://raw.githubusercontent.com/haochen2049/vllm-headroom/main/assets/report.png)](https://haochen2049.github.io/vllm-headroom/)
+
+> Status: alpha (0.1). Tested against vLLM 0.31 on one GPU type; see [Validation](#validation).
+> Not affiliated with the vLLM project.
+
+## Why
+
+Dashboards show what already happened, and forecasting tools extrapolate it. Neither tells you what
+happens after you change something, such as more traffic, a different `--max-num-seqs`, one fewer
+replica or a longer prompt mix, or how close you are to the point where latency falls off a cliff.
+Load testing every option is slow. A model fitted to *your* server answers these questions in seconds.
+
+## Quick start
+
+```bash
+pip install vllm-headroom
+
+# no GPU? profile a bundled simulated server and open the sample report (about 2-3 minutes)
+vllm-headroom demo -o demo_report.html
+```
+
+Against a real server, ideally an idle replica or a staging copy:
+
+```bash
+# 1. profile (about 3-4 minutes); pass the server's own settings if you know them
+vllm-headroom profile --url http://localhost:8000 \
+    --max-num-seqs 256 --max-num-batched-tokens 8192 -o profile.json
+
+# 2. analyse against your SLO and current traffic
+vllm-headroom analyze profile.json --slo ttft_p95=2 --slo tpot_p95=0.1 \
+    --current-rate 12 --workload my_requests.jsonl -o headroom.html
+
+# 3. compare configurations quickly
+vllm-headroom whatif profile.json --slo ttft_p95=2 \
+    --change max_num_seqs=512 --change replicas=3 --change kv_scale=0.5
+```
+
+`--workload` is a JSONL file with one request per line, `{"prompt_tokens": 812, "output_tokens": 230}`
+(OpenAI `usage` objects also work). Without it, `--prompt-mean/--output-mean` give a log-normal mix,
+and the default is the mix used during profiling. SLO metrics are `ttft` (time to first token),
+`tpot` (time per output token after the first) and `e2e` (end-to-end), each with any percentile:
+`ttft_p95=2`, `e2e_p99=30`, `tpot_p50=0.05`.
+
+## How it works
+
+**1. Profile.** Four phases against the OpenAI-compatible endpoint:
+
+| phase | what is sent | what it identifies |
+|---|---|---|
+| prefill sweep | single requests, prompts from 128 to 8192 tokens, idle server | prefill cost per token and fixed overhead |
+| decode waves | `c` identical requests at once, `c` = 1, 2, 4 ... 256, short and long prompts | step time versus batch size and context length |
+| admission probes | short requests sent while a wave decodes | how many steps a new request waits before it is scheduled |
+| validation | open-loop Poisson load at 30, 50 and 70% of predicted saturation | nothing: held out for the backtest |
+
+Prompts are random token ids, so prefix caching cannot shorten prefill and lengths are exact;
+`ignore_eos` fixes output lengths. Streaming chunk timestamps give inter-token gaps; only gaps with no
+prefill from another request are used for decode fitting. KV capacity is read from `/metrics`, and
+`max_num_seqs` is inferred from the waves if you do not pass it.
+
+**2. Fit.** Engine step time is modelled as
+
+```
+step = b0 + b_dec·(decoding sequences) + b_ctx·(their context tokens)
+          + b_pf·(prefill tokens) + b_pfatt·(prefill attention term)
+```
+
+fitted with robust least squares, plus a client-visible `overhead` and an `admit_lag` (vLLM V1 plans
+the next step while the current one runs, so a new request waits about one to two extra steps).
+A bootstrap over profiling requests gives parameter uncertainty.
+
+**3. Simulate.** A discrete-event simulator of vLLM-style continuous batching: chunked prefill under
+`max_num_batched_tokens`, `max_num_seqs`, KV capacity with recompute preemption, several replicas
+behind a least-loaded or round-robin router. Decode-only stretches are advanced in closed form, so a
+1,500-request run takes well under a second.
+
+**4. Report.** The maximum Poisson rate meeting every SLO (bisection), the saturation rate, latency
+curves with a 10–90% range across bootstrap refits, what-if comparisons and a backtest against the
+held-out validation load.
+
+## Validation
+
+One run so far: Qwen2.5-7B-Instruct on vLLM 0.31.0, one NVIDIA RTX PRO 6000 Blackwell (96 GB),
+`--max-num-seqs 64 --max-num-batched-tokens 4096 --gpu-memory-utilization 0.5`. The profile, report
+and check script are in [`examples/qwen2.5-7b-vllm0.31`](https://github.com/haochen2049/vllm-headroom/tree/main/examples/qwen2.5-7b-vllm0.31). The workload
+is the default log-normal mix (prompt mean 505, output mean 130 tokens). `max_num_seqs` was not passed
+and was inferred correctly (64).
+
+**Backtest** (held-out validation load from the same profiling run, never used for fitting):
+
+| load (of predicted saturation) | requests | TTFT p50 | TTFT p95 | E2E p50 | E2E p95 |
+|---|---|---|---|---|---|
+| 30% (6.5 req/s) | 161 | +8% | +7% | +2% | +2% |
+| 50% (10.8 req/s) | 268 | +9% | +12% | +2% | +2% |
+| 70% (15.1 req/s) | 376 | +8% | +14% | +3% | +4% |
+
+**Near saturation** (a separate later run against the same server with the fit frozen,
+`near_saturation_check.py`):
+
+| load | requests | TTFT p50 | TTFT p95 | E2E p50 | E2E p95 |
+|---|---|---|---|---|---|
+| 85% (18.5 req/s) | 833 | +7% | +5% | +3% | +5% |
+| 95% (20.7 req/s) | 931 | 0% | −19% | −3% | 0% |
+
+Entries are prediction error relative to measurement (positive = predicted slower). End-to-end
+latency is within 5% everywhere. TTFT is predicted somewhat pessimistically up to 85% load. At 95%
+load the TTFT tail is under-predicted by 19%, because close to saturation a small error in the rate
+moves the tail a lot. Treat the last few percent before saturation as unreliable and keep a margin.
+This is one model, one GPU and one workload; results on other setups may differ. If you run it on
+other hardware, please [share your backtest](https://github.com/haochen2049/vllm-headroom/issues/new?template=backtest.yml): it is the most useful contribution right now.
+
+The unit tests also check the simulator against a plain step-by-step reference implementation
+(identical results), recover known parameters from synthetic data, and run the full
+profile → fit → report pipeline over HTTP against the bundled simulated server.
+
+## Limits
+
+- One replica is profiled; multiple replicas are modelled as identical copies.
+- Arrivals are Poisson at a constant rate. Bursty traffic needs more headroom than reported.
+- Not modelled: prefix-cache hits, speculative decoding, LoRA adapter switching, multi-model serving,
+  disaggregated prefill/decode, pipeline-parallel bubbles, network between client and router.
+- Batch sizes and prompt lengths beyond what was profiled are extrapolated.
+- The uncertainty range reflects profiling noise, not every way production can differ.
+- Profiling sends real load. Run it on a replica that is not serving users.
+
+## Output files
+
+`profile` writes raw timings (no prompt or output text). `analyze` writes `headroom.html` and
+`headroom.json` with the fitted parameters, curves, headroom, backtest and what-if results.
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+pytest -q
+vllm-headroom mock-server --port 8001   # a simulated OpenAI-compatible server for experiments
+```
+
+## License
+
+Apache-2.0

vllm_headroom-0.1.0/pyproject.toml

@@ -0,0 +1,48 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "vllm-headroom"
+dynamic = ["version"]
+description = "Estimate how much more load a vLLM deployment can take before it breaks its latency SLO"
+readme = "README.md"
+license = {text = "Apache-2.0"}
+requires-python = ">=3.10"
+dependencies = ["numpy>=1.24", "httpx>=0.25"]
+keywords = ["vllm", "llm", "inference", "capacity-planning", "slo", "queueing", "simulation"]
+classifiers = [
+  "Development Status :: 3 - Alpha",
+  "Environment :: Console",
+  "Intended Audience :: Developers",
+  "Intended Audience :: System Administrators",
+  "License :: OSI Approved :: Apache Software License",
+  "Operating System :: OS Independent",
+  "Programming Language :: Python :: 3",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.11",
+  "Programming Language :: Python :: 3.12",
+  "Topic :: Scientific/Engineering :: Artificial Intelligence",
+  "Topic :: System :: Benchmark",
+]
+
+[project.urls]
+Homepage = "https://github.com/haochen2049/vllm-headroom"
+Source = "https://github.com/haochen2049/vllm-headroom"
+Issues = "https://github.com/haochen2049/vllm-headroom/issues"
+"Example report" = "https://haochen2049.github.io/vllm-headroom/"
+
+[project.optional-dependencies]
+dev = ["pytest>=7"]
+
+[project.scripts]
+vllm-headroom = "vllm_headroom.cli:main"
+
+[tool.setuptools.dynamic]
+version = {attr = "vllm_headroom.__version__"}
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

vllm_headroom-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

vllm_headroom-0.1.0/src/vllm_headroom/__init__.py

@@ -0,0 +1,3 @@
+"""Estimate how much more load a vLLM deployment can take before it breaks its latency SLO."""
+
+__version__ = "0.1.0"