[doc] Update perf_overview.md for release 0.21 #6270
Conversation
WalkthroughThe documentation for performance benchmarking was updated to include a new NVIDIA blog reference, revised throughput tables with updated GPU models and tensor parallel sizes, and adjusted performance values. Tables were reformatted for compactness, notes added about Llama 4 performance issues on short sequences, version references updated, and minor clarifications made to benchmark details. Changes
Estimated code review effort🎯 1 (Trivial) | ⏱️ ~4 minutes Possibly related PRs
Suggested labels
Suggested reviewers
Poem
Note ⚡️ Unit Test Generation is now available in beta!Learn more here, or try it out under "Finishing Touches" below. ✨ Finishing Touches🧪 Generate unit tests
🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (Invoked using PR comments)
Other keywords and placeholders
Documentation and Community
|
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
zbpatel
force-pushed
the
user/zpatel/perf_overview_0.21
branch
from
4052b26 to
b8d1a6d
Compare
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (8)
docs/source/performance/perf-overview.md (8)
34-38: Table header is hard to parse – consider collapsing the multi-row headerThe current five rows of “empty” cells render as a single table with several nearly-blank rows, which looks broken in most Markdown renderers and makes copy-pasting awkward. A single-row header is clearer:
-| | GPU: | B200 | GB200 180GB | -|:-----------------------------|:---|:----------|:--------------| -| | TP Size | 1 | 1 | -| ISL, OSL | | | | -| | | | | +| ISL | OSL | TP Size | B200 | GB200 180GB | +|:--:|:--:|:------:|------:|------------:|This keeps the same information while eliminating the “blank” rows.
53-57: Same multi-row-header issue as aboveThe 405 B FP4 table repeats the pattern that causes rendering problems. Consider applying the same single-row-header refactor for consistency.
80-96: Large 8 B FP8 table: readability could be improvedWith three GPUs in one table, a short caption or footnote explaining why all TP sizes are “1” would help readers unfamiliar with the context.
99-115: 70 B FP8 table header spacingMinor: “GPU:” is only useful if there is a second descriptor. Dropping the colon keeps the heading tighter:
-| | GPU: | H100 80GB HBM3 | H200 141GB HBM3 | +| | GPU | H100 80GB HBM3 | H200 141GB HBM3 |
117-133: Lint failure (MD058) – add a blank line before the table
markdownlintflags this table because it is not surrounded by blank lines.-#### Llama 3.1 405B FP8 -| | GPU: | H100 80GB HBM3 | H200 141GB HBM3 | +#### Llama 3.1 405B FP8 + +| | GPU | H100 80GB HBM3 | H200 141GB HBM3 |
136-136: Style: turn the long sentence into a blockquote or note for better emphasis-*Performance for Llama 4 on sequence lengths less than 8,192 tokens is affected by an issue introduced in v0.21. To reproduce the Llama 4 performance noted here, please use v0.20 +> **NOTE:** An issue introduced in v0.21 affects Llama 4 performance for sequence lengths **< 8 192** tokens. To reproduce the numbers below, run **v0.20**.
227-249: YAML options block: shortcode for long listsShowing the full
cuda_graph_batch_sizeslist clutters the doc. Consider collapsing with an ellipsis to keep focus on the relevant flags:use_cuda_graph: true cuda_graph_padding_enabled: true cuda_graph_batch_sizes: [1, 2, 4, 8, 16, … , 4096, 8192]This still conveys the idea while shortening ~20 lines.
10-12: Typo: “situaiton” → “situation”-Tuning batch sizes, parallelism configurations, and other options may lead to improved performance depending on your situaiton. +Tuning batch sizes, parallelism configurations, and other options may lead to improved performance depending on your situation.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(4 hunks)
🧰 Additional context used
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
117-117: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (1)
docs/source/performance/perf-overview.md (1)
15-16: Blog link addition looks goodThe cross-reference to FrankD412’s benchmarking post is a valuable pointer for readers who want deeper guidance. Nothing to change here.
zbpatel
changed the title
|
I just pushed the tables with v0.21 numbers, there are 2 minor TODOs left:
|
Ok turns out I was wrong about 2, we kept these values the same. |
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
There was a problem hiding this comment.
Actionable comments posted: 0
♻️ Duplicate comments (2)
docs/source/performance/perf-overview.md (2)
34-49: Missing * marker requested in prior review
A previous reviewer (see “laikhtewari” comment) asked for an asterisk next to every updated number. These FP4-70B figures were changed but still lack the marker.
53-68: Same asterisk request applies here
The 405B-FP4 table is also updated and still omits the requested asterisks.
🧹 Nitpick comments (4)
docs/source/performance/perf-overview.md (4)
80-96: Insert a blank line before the table to satisfy MD058
markdownlint flags this because the table immediately follows the heading.-#### Llama 3.1 8B FP8 -| | GPU: | GH200 | H100 80GB HBM3 | H200 141GB HBM3 | +#### Llama 3.1 8B FP8 + +| | GPU: | GH200 | H100 80GB HBM3 | H200 141GB HBM3 |
99-115: Blank line needed before this table as well
Same MD058 issue as above.-#### Llama 3.3 70B FP8 -| | GPU: | H100 80GB HBM3 | H200 141GB HBM3 | +#### Llama 3.3 70B FP8 + +| | GPU: | H100 80GB HBM3 | H200 141GB HBM3 |
117-133: Blank line missing before FP8-405B table
This is the specific line markdownlint reported (MD058).-#### Llama 3.1 405B FP8 -| | GPU: | H100 80GB HBM3 | H200 141GB HBM3 | +#### Llama 3.1 405B FP8 + +| | GPU: | H100 80GB HBM3 | H200 141GB HBM3 |
136-138: Add blank line between the note and the table
Keeps consistency and resolves MD058 for this section.-Note: Performance for Llama 4 on sequence lengths less than 8,192 tokens is affected by an issue introduced in v0.21. To reproduce the Llama 4 performance noted here, please use v0.20 -| | GPU | H200 141GB HBM3 | H100 80GB HBM3 | +Note: Performance for Llama 4 on sequence lengths less than 8,192 tokens is affected by an issue introduced in v0.21. To reproduce the Llama 4 performance noted here, please use v0.20 + +| | GPU | H200 141GB HBM3 | H100 80GB HBM3 |
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(4 hunks)
🧰 Additional context used
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
117-117: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (2)
docs/source/performance/perf-overview.md (2)
15-16: Nice contextual link added
The extra reference to thetrtllm-benchblog post is helpful for users who want deeper tuning guidance.
227-249: Appreciate the concretellm_options.ymlexample
Including the exact CUDA-graph batch-size list makes reproduction straightforward.
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
docs/source/performance/perf-overview.md (1)
10-12: Fix typo in introductory paragraph
situaiton➜situation
Minor, but it appears early in the doc and slightly detracts from polish.-Tuning batch sizes, parallelism configurations, and other options may lead to improved performance depending on your situaiton. +Tuning batch sizes, parallelism configurations, and other options may lead to improved performance depending on your situation.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(4 hunks)
🧰 Additional context used
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
117-117: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
|
@laikhtewari do you mind reviewing / approving changes? I think the PR is blocked waiting for an approval from you. |
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (1)
docs/source/performance/perf-overview.md (1)
125-142: Surround the 405B-FP8 table with blank lines (MD058)The table starting on Line 126 is not preceded and followed by blank lines, causing the
MD058 blanks-around-tableslinter failure.-#### Llama 3.1 405B FP8 -| | GPU: | H100 | H200 | +#### Llama 3.1 405B FP8 + +| | GPU: | H100 | H200 | @@ -| 20000, 2000 | | 363.27 | 509.87 | +| 20000, 2000 | | 363.27 | 509.87 | +
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(6 hunks)
🧰 Additional context used
🧠 Learnings (1)
docs/source/performance/perf-overview.md (1)
Learnt from: moraxu
PR: #6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
26-26: Trailing punctuation in heading
Punctuation: ':'
(MD026, no-trailing-punctuation)
126-126: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (1)
docs/source/performance/perf-overview.md (1)
126-133: Insert a blank line before the table to satisfy MD058The table following “#### Llama 3.1 405B FP8” is not preceded by a blank line, triggering MD058.
-#### Llama 3.1 405B FP8 -| | GPU: | H100 | H200 | +#### Llama 3.1 405B FP8 + +| | GPU: | H100 | H200 |No other adjustments are needed—there’s already a trailing blank line after the table.
🧹 Nitpick comments (1)
docs/source/performance/perf-overview.md (1)
26-35: Clarify GH200 memory figures to avoid reader confusionLine 30 mixes GPU memory (96 GB HBM3) with system memory (480 GB LPDDR5X) in a single parenthetical, which previously generated questions.
Consider making the split explicit (e.g. “GH200 96 GB HBM3 GPU / 480 GB LPDDR5X system”) to prevent misinterpretation.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(5 hunks)
🧰 Additional context used
🧠 Learnings (1)
docs/source/performance/perf-overview.md (1)
Learnt from: moraxu
PR: #6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
37-37: Fenced code blocks should have a language specified
(MD040, fenced-code-language)
127-127: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (1)
docs/source/performance/perf-overview.md (1)
15-16: Blog-post reference looks goodAdds useful context for users wanting deeper benchmarking guidance.
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (1)
docs/source/performance/perf-overview.md (1)
128-145: Surround the table with blank lines to satisfy MD058Markdown-lint requires a blank line before and after tables. Insert one above line 129 and one after line 144.
-#### Llama 3.1 405B FP8 -| | GPU: | H100 | H200 | +#### Llama 3.1 405B FP8 + +| | GPU: | H100 | H200 | ... -| 20000, 2000 | | 363.27 | 509.87 | + +| 20000, 2000 | | 363.27 | 509.87 | +This unblocks CI and improves readability.
🧹 Nitpick comments (1)
docs/source/performance/perf-overview.md (1)
26-35: Clarify GH200 memory spec to avoid renewed confusionThe parenthetical “(480GB LPDDR5X)” re-introduces the same ambiguity that was removed in an earlier commit (see Juney’s prior comment). Readers may again assume 480 GB is GPU HBM rather than system RAM.
- - GH200 96GB HBM3 (480GB LPDDR5X) + - GH200 96 GB HBM3 (Grace-CPU 480 GB LPDDR5X system RAM)Adding a brief note or splitting GPU vs. system memory explicitly will prevent misinterpretation.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(5 hunks)
🧰 Additional context used
🧠 Learnings (1)
docs/source/performance/perf-overview.md (1)
Learnt from: moraxu
PR: #6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
38-38: Fenced code blocks should have a language specified
(MD040, fenced-code-language)
129-129: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (1)
docs/source/performance/perf-overview.md (1)
15-16: 👍 Good contextual reference addedThe link to the NVIDIA blog post gives readers a clear next step for deeper benchmarking guidance. No further action needed here.
zbpatel
added
the
Doc
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
docs/source/performance/perf-overview.md (1)
15-16: Nice contextual reference – tiny readability polish suggestedThe added blog link is great for readers who want deeper benchmarking guidance.
Consider inserting a comma after “trtllm-bench” so the sentence flows a bit better:-For more information on benchmarking with `trtllm-bench` see this NVIDIA +For more information on benchmarking with `trtllm-bench`, see this NVIDIA
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
docs/source/performance/perf-overview.md(5 hunks)
🧰 Additional context used
🧠 Learnings (1)
docs/source/performance/perf-overview.md (1)
Learnt from: moraxu
PR: #6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
🪛 markdownlint-cli2 (0.17.2)
docs/source/performance/perf-overview.md
129-129: Tables should be surrounded by blank lines
(MD058, blanks-around-tables)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (1)
docs/source/performance/perf-overview.md (1)
26-32: GH200 bullet re-introduces the 480 GB figure that previously caused confusionPast discussion removed the “480 GB” notation because it mixes host LPDDR5X capacity
with GPU HBM capacity and led to mis-interpretation (see Juney’s prior comment).
The new bullet adds it back:- GH200 96GB HBM3 (480GB LPDDR5X)Unless the host-memory number is essential here (it is not referenced anywhere in
the tables), please drop the parenthetical to stay consistent with earlier fixes
and the other bullets:- - GH200 96GB HBM3 (480GB LPDDR5X) + - GH200 96GB HBM3
|
/bot skip --comment "Documentation update." |
|
PR_Github #13575 [ skip ] triggered by Bot |
|
PR_Github #13575 [ skip ] completed with state |
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]>
Signed-off-by: zpatel <[email protected]> Signed-off-by: Lanyu Liao <[email protected]>
Signed-off-by: zpatel <[email protected]>
Summary by CodeRabbit
Description
This MR updates the
perf_overview.mddocument with benchmarks from the 0.21 release.In addition to updating the numbers, I added a link to @FrankD412 's blog post on benchmarking with trtllm-bench
Test Coverage
N/A doc
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.
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.