[V1][Perf] Faster incremental detokenization

[njhill]

Exploit tokenizers library's new DecodeStream functionality for "fast" HF tokenizers.

This won't help with mistral tokenizers unfortunately.

cc @robertgshaw2-redhat

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[WoosukKwon]

@njhill Thanks for doing this! Do you have any performance numbers by any chance?

[njhill]

@WoosukKwon looks like it's noticeable for higher QPS:

Before:

============ Serving Benchmark Result ============
Successful requests:                     6000      
Benchmark duration (s):                  81.47     
Total input tokens:                      1322102   
Total generated tokens:                  1205048   
Request throughput (req/s):              73.65     
Output token throughput (tok/s):         14791.01  
Total Token throughput (tok/s):          31018.76  
---------------Time to First Token----------------
Mean TTFT (ms):                          207.24    
Median TTFT (ms):                        179.43    
P99 TTFT (ms):                           1081.34   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          22.78     
Median TPOT (ms):                        22.36     
P99 TPOT (ms):                           34.43     
---------------Inter-token Latency----------------
Mean ITL (ms):                           26.03     
Median ITL (ms):                         25.08     
P99 ITL (ms):                            48.39     
==================================================

After:

============ Serving Benchmark Result ============
Successful requests:                     6000      
Benchmark duration (s):                  80.39     
Total input tokens:                      1322102   
Total generated tokens:                  1205048   
Request throughput (req/s):              74.64     
Output token throughput (tok/s):         14990.72  
Total Token throughput (tok/s):          31437.59  
---------------Time to First Token----------------
Mean TTFT (ms):                          166.37    
Median TTFT (ms):                        136.98    
P99 TTFT (ms):                           1054.43   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          22.49     
Median TPOT (ms):                        22.21     
P99 TPOT (ms):                           34.91     
---------------Inter-token Latency----------------
Mean ITL (ms):                           23.49     
Median ITL (ms):                         21.41     
P99 ITL (ms):                            44.67     
==================================================

[WoosukKwon]

@njhill The speedup looks promising! 🚀

Is this PR ready for review? If so, could you please add tests?

[njhill]

@WoosukKwon yes it's ready apart from some additional tests.

[WoosukKwon]

@njhill Got it. Could you please resolve the merge conflict?

[njhill]

@WoosukKwon I'm just updating tests, will push those and resolve conflict at the same time.

[njhill]

Ready now @WoosukKwon, sorry the tests turned out to be a bit of a can of worms 😅

[WoosukKwon]

@robertgshaw2-redhat would you be interested in reviewing this PR? I'm not sure if any reviewer is assigned.

[njhill]

I will push a fix for the remaining test failure asap.

[robertgshaw2-redhat]

nice way to skip the priming.

[robertgshaw2-redhat]

Nice work.

[njhill]

Looks like the remaining test failures are "real", need to dig into the behavior differences that are causing them.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @njhill.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[njhill]

Addressed the remaining test failures (due to incorrect behaviour of the existing incremental detokenization which doesn't skip some special tokens when it should). So should now be good to merge.