[None] [doc] Add Mixed Precision Context and Generation section to Disagg #8164
Conversation
examples/disaggregated/README.md
Outdated
|
|
||
| ## Mixed Precision Context and Generation | ||
|
|
||
| In disaggregated serving, the context (prefill) workers and generation (decode) workers have different performance characteristics: prefill workers are more compute-bound while decode workers are more memory-bound. Therefore, it may be beneficial to run prefill workers in higher precision. Running these workers with different precisions also enables the ability to interpolate between performance/compute trade-offs of different quantization levels. |
There was a problem hiding this comment.
No need to specify prefill/decode, the terms are not used in the rest of the document.
examples/disaggregated/README.md
Outdated
|
|
||
| ## Mixed Precision Context and Generation | ||
|
|
||
| In disaggregated serving, the context (prefill) workers and generation (decode) workers have different performance characteristics: prefill workers are more compute-bound while decode workers are more memory-bound. Therefore, it may be beneficial to run prefill workers in higher precision. Running these workers with different precisions also enables the ability to interpolate between performance/compute trade-offs of different quantization levels. |
There was a problem hiding this comment.
prefill workers are more compute-bound while decode workers are more memory-bound.
Use context / generation as well here. Remove 'more'
juney-nvidia
added
Community want to contribute
|
The full results for MMLU, GSM8K tests using 16/16, 16/8, 8/8, 16/4, 4/4, and 4/16 precisions for Llama 3.1 8B are available here: https://docs.google.com/spreadsheets/d/1jyG4WFsrPSoBISw72wioJjgLJZL2Yqe7uRHoFIpyfcE/edit?usp=sharing |
timothygao8710
requested review from
Shixiaowei02,
Tabrizian,
chuangz0 and
chzblych
📝 WalkthroughWalkthroughAdded documentation section to disaggregated README describing mixed-precision serving, explaining compute vs. memory-bound worker trade-offs, and providing prerequisites and example workflows for quantized checkpoints with identical KV cache dtypes. Duplicate content was added. Changes
Estimated code review effort🎯 2 (Simple) | ⏱️ ~10 minutes Rationale: Documentation-only changes with straightforward additions. Review involves verifying content accuracy, checking for the noted duplication, and ensuring command snippets are correct. No logic or structural code changes to evaluate. Pre-merge checks and finishing touches✅ Passed checks (3 passed)
✨ Finishing touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
examples/disaggregated/README.md (1)
203-260: Consider consolidating duplication with the Dynamic scaling section.The new "Mixed Precision Context and Generation" section overlaps significantly with the "Dynamic scaling" section below (lines 261–286) in terms of the workflow and command patterns shown. Both sections demonstrate:
- Starting context servers with
--server_role CONTEXTand--metadata_server_config_file- Starting generation servers with
--server_role GENERATIONand--metadata_server_config_file- Launching the disaggregated server with
-m ./metadata_config.yamlThis duplication could confuse readers about whether the features are prerequisites for each other or independent capabilities. Consider clarifying the relationship or consolidating the examples.
Review the overlap between lines 235–253 and lines 273–279 to determine if one example can reference the other or if the sections should be restructured.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
examples/disaggregated/README.md(1 hunks)
🧰 Additional context used
🪛 LanguageTool
examples/disaggregated/README.md
[grammar] ~209-~209: There might be a mistake here.
Context: ...le mixed precision serving, you'll need: 1. A quantized checkpoint created with [Ten...
(QB_NEW_EN)
[grammar] ~210-~210: There might be a mistake here.
Context: ...t created with TensorRT Model Optimizer 2. The original unquantized checkpoint 3. B...
(QB_NEW_EN)
[grammar] ~211-~211: There might be a mistake here.
Context: ...) 2. The original unquantized checkpoint 3. Both checkpoints must use the same KV ca...
(QB_NEW_EN)
🔇 Additional comments (1)
examples/disaggregated/README.md (1)
205-205: Verify past feedback has been addressed.The text correctly uses "context workers" and "generation workers" terminology (not prefill/decode) and omits the "more" modifier ("compute-bound" not "more compute-bound"), which aligns with feedback from the previous review round. ✓
examples/disaggregated/README.md
Outdated
|
|
||
| ### Example (BF 16 Gen, FP 8 Ctx) | ||
|
|
||
| A quantized checkpoint can be created `--kv_cache_qformat none`. |
There was a problem hiding this comment.
Complete the sentence at line 216.
The sentence is incomplete: "A quantized checkpoint can be created --kv_cache_qformat none." is missing a preposition or connector between the verb and the flag.
-A quantized checkpoint can be created `--kv_cache_qformat none`.
+A quantized checkpoint can be created with `--kv_cache_qformat none`.📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| A quantized checkpoint can be created `--kv_cache_qformat none`. | |
| A quantized checkpoint can be created with `--kv_cache_qformat none`. |
🤖 Prompt for AI Agents
In examples/disaggregated/README.md around line 216, the sentence "A quantized
checkpoint can be created `--kv_cache_qformat none`." is missing a connector;
update the line to include a preposition such as "with" or "using" (for example:
"A quantized checkpoint can be created with `--kv_cache_qformat none`.") to make
the sentence grammatically correct and clear.
3 participants
Summary by CodeRabbit
Description
This PR updates the Disaggregated Serving README to document how to run context and generation servers in mixed-precision using model opt. An example usage for BF16 prefill + FP8 decode is included.
Test Coverage
Performance and accuracy tests were run for FP8 and NVFP4 with llama 3.1 8B base model on GSM8k and MMLU, with 1 Ctx/Gen Server. Example:
FP16/16 (2xH100):
"results": { "gsm8k": { "alias": "gsm8k", "exact_match,strict-match": 0.7558756633813495, "exact_match_stderr,strict-match": 0.011832404674077592, "exact_match,flexible-extract": 0.7839272175890827, "exact_match_stderr,flexible-extract": 0.011336531489638858 } },"start_time": 34733.35253426, "end_time": 35711.569743744, "total_evaluation_time_seconds": "978.2172094840062"FP16/8 (2xH100):
"results": { "gsm8k": { "alias": "gsm8k", "exact_match,strict-match": 0.7354056103108415, "exact_match_stderr,strict-match": 0.012150554001563231, "exact_match,flexible-extract": 0.7626990144048522, "exact_match_stderr,flexible-extract": 0.011718409178739442 } },"start_time": 33387.524966359, "end_time": 34123.906708314, "total_evaluation_time_seconds": "736.3817419550032"PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
[ x ] Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
Details
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.