[TPU][V1] Enable Top-P

[NickLucche]

Follow up to #15489.
This PR enables the work done #15736.
We split up the contribs to better monitor performance as described more in detail here #16268.

Here's the benchmark I ran on single chip with and without top-p:

Benchmark commands

MODEL=meta-llama/Llama-3.1-8B-Instruct
# Server
VLLM_XLA_CHECK_RECOMPILATION=1 VLLM_USE_V1=1 vllm serve $MODEL \
 --disable-log-requests \
 --port 8004 \
 --gpu-memory-utilization 0.95 \
 --max-num-seqs 512 \
 --max-num-batched-tokens 512 \
 --tensor-parallel-size 1 \
 --max-model-len 2048 > "$VLLM_LOG" 2>&1 &
# Client with the newly added --top-k --top-p options
python benchmarks/benchmark_serving.py \
    --backend vllm \
    --model $MODEL \
    --dataset-name random \
    --random-input-len 1720 \
    --random-output-len 128 \
    --random-prefix-len 0 \
    --ignore-eos \
    --top-k 8 --top-p 0.5 \
    --port 8004 > "$BM_LOG"

# with top-p (this PR) 

 ============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  128.28    
Total input tokens:                      1720000   
Total generated tokens:                  128000    
Request throughput (req/s):              7.80      
Output token throughput (tok/s):         997.85    
Total Token throughput (tok/s):          14406.52  
---------------Time to First Token----------------
Mean TTFT (ms):                          63332.60  
Median TTFT (ms):                        62792.18  
P99 TTFT (ms):                           124350.61 
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          32.65     
Median TPOT (ms):                        33.30     
P99 TPOT (ms):                           63.55     
---------------Inter-token Latency----------------
Mean ITL (ms):                           37.65     
Median ITL (ms):                         33.21     
P99 ITL (ms):                            34.56     
==================================================

# without top-p (26507f897) 

============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  129.28    
Total input tokens:                      1720000   
Total generated tokens:                  128000    
Request throughput (req/s):              7.74      
Output token throughput (tok/s):         990.13    
Total Token throughput (tok/s):          14295.02  
---------------Time to First Token----------------
Mean TTFT (ms):                          64067.78  
Median TTFT (ms):                        63521.31  
P99 TTFT (ms):                           125308.93 
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          32.56     
Median TPOT (ms):                        33.42     
P99 TPOT (ms):                           63.77     
---------------Inter-token Latency----------------
Mean ITL (ms):                           37.30     
Median ITL (ms):                         33.33     
P99 ITL (ms):                            34.65     
==================================================

Performance is looking good here. We should probably run a larger sweep just to be safe.

PS: performance with/without topk/p sampling-params (eg setting only temperature=0.2) is virtually identical as the graph being executed is the same. This is not the case when all requests in the batch have temperature=0 (all_greedy).

cc @hyeygit

[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.

🚀

[mgoin]

Thanks, LGTM

[DarkLight1337]

Can you merge from main to fix CI?

[yaochengji]

LGTM, thanks for the contribution!