[Metrics] Deprecate TPOT in favor of ITL #24110
Merged
+64
−26
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
The only case where we don't want to assert the existance of a metric is where it is deprecated and we're not showing hidden deprecated metrics. Signed-off-by: Mark McLoughlin <[email protected]>
markmc
requested review from
DarkLight1337,
WoosukKwon,
aarnphm,
alexm-redhat,
comaniac,
njhill,
robertgshaw2-redhat,
simon-mo,
youkaichao,
ywang96 and
zhuohan123
as code owners
5 tasks
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request correctly deprecates the vllm:time_per_output_token_seconds (TPOT) metric in favor of the more accurately named vllm:inter_token_latency_seconds (ITL). The changes are consistently applied across the codebase, including metrics definitions, logging, tests, and the Grafana dashboard example. The deprecation strategy of retaining the old metric for backward compatibility while introducing the new one is sound. I've found one minor issue with the documentation of the new metric, which appears to be a copy-paste error.
As per vllm-project#24015, what we currently call as TPOT should instead be called ITL since what we are actually measuring is the time between iterations, and a single iteration can produce multiple tokens. Signed-off-by: Mark McLoughlin <[email protected]>
markmc
force-pushed
the
metrics-rename-tpot-to-itl
branch
from
b176439 to
09dbc43
Compare
DarkLight1337
approved these changes
Sep 2, 2025
github-actions
bot
added
the
ready
845473182
pushed a commit
to 845473182/vllm
that referenced
this pull request
Sep 3, 2025
* 'main' of https://github.com/845473182/vllm: (457 commits) [BugFix] Fix routed_scaling_factor double mul for dots1 and glm4 MoE models (vllm-project#24132) [Misc] Add check for dual_chunk_attention (vllm-project#24070) [Doc]: fix typos in Python comments (vllm-project#24115) [Doc]: fix typos in Python comments (vllm-project#24093) [Compile] Fix Compile Warning for `w4a8_mm_entry.cu` (vllm-project#23660) fix some typos (vllm-project#24071) [V1] Wrapper which plumbs request-level logits processors into vLLM batch-level logits processing (vllm-project#23656) Upgrade xgrammar to 0.1.23 (vllm-project#22988) Update release pipeline post PyTorch 2.8.0 update (vllm-project#24073) [XPU] Fix the bug of LoRA logits on the XPU platform (vllm-project#24081) [CI/Build] Disable SiluMul NVFP4 quant fusion tests (vllm-project#24121) [Bug] R1 Accuracy: Fix `routed_scaling_factor` Double Mul Issue (vllm-project#24119) [AMD][Kernel][Bugfix] Cast offsets tensor bn to tl.int64 to avoid GPU segfault (vllm-project#23692) [CI] Enable all hf transformers baselines in test_hybrid (vllm-project#23936) [Log] Only Print Profiler Results on Rank 0 (vllm-project#23370) Fix weights loading for Apertus (vllm-project#24100) [Metrics] Deprecate TPOT in favor of ITL (vllm-project#24110) [Bugfix] Fix packed_factor missing attribute error (vllm-project#23902) Run ruff format on a few files. (vllm-project#24075) [Bugfix] Fix transform_config parsing in Compressed Tensors (vllm-project#23945) ...
eicherseiji
pushed a commit
to eicherseiji/vllm
that referenced
this pull request
Sep 9, 2025
Signed-off-by: Mark McLoughlin <[email protected]>
FeiDaLI
pushed a commit
to FeiDaLI/vllm
that referenced
this pull request
Sep 25, 2025
Signed-off-by: Mark McLoughlin <[email protected]>
1 task
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
As per #24015, what we currently call as TPOT should instead be called ITL since what we are actually measuring is the time between iterations, and a single iteration can produce multiple tokens.
I'm flagging the TPOT metric as deprecated from 0.11 - even if this gets released on a 0.10.x release, I think we should only start the deprecation period from when it gets released in a new minor 0.N.0 release.