[GDN] Fused all preprocessing into one kernel

[apinge]

Motivation

Same change as vllm-project/vllm#38787

Modifications

Accuracy Tests

Qwen3.5-27B-PTPC-compressor

Tested on MI308X

lm_eval --model local-completions \
    --model_args '{"base_url": "http://localhost:9080/v1/completions", "model": "/models/Qwen3.5-27B-PTPC-compressor", "num_concurrent": 256, "max_retries": 10, "max_gen_toks": 2048}' \
    --tasks gsm8k \
    --batch_size auto \
    --num_fewshot 5 \
    --trust_remote_code 2>&1 | tee gsm8k_eval.log

local-completions ({'base_url': 'http://localhost:9080/v1/completions', 'model': '/models/Qwen3.5-27B-PTPC-compressor', 'num_concurrent': 256, 'max_retries': 10, 'max_gen_toks': 2048}), gen_kwargs: ({}), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7597|±  |0.0118|
|     |       |strict-match    |     5|exact_match|↑  |0.8537|±  |0.0097|

qwen3p5_397B_ptpc

RESULT_FILE: accuracy_test_results/offical_qwen3p5_397B_ptpc_gsm8k_results.json
GSM8K flexible extract score: 0.95

Benchmarking and Profiling

Tested on MI308X

./run_pure_text_bench.sh   
bench model: /models/Qwen3.5-27B-PTPC-compressor/
input tokens: 8000
output tokens: 500
max concurrency: 1
num prompts: 32
dataset-name: random


#Input tokens: 256000
#Output tokens: 16000

============ Serving Benchmark Result ============
Backend:                                 sglang    
Traffic request rate:                    inf       
Max request concurrency:                 1         
Successful requests:                     32        
Benchmark duration (s):                  265.83    
Total input tokens:                      256000    
Total input text tokens:                 256000    
Total generated tokens:                  16000     
Total generated tokens (retokenized):    15994     
Request throughput (req/s):              0.12      
Input token throughput (tok/s):          963.03    
Output token throughput (tok/s):         60.19     
Peak output token throughput (tok/s):    71.00     
Peak concurrent requests:                2         
Total token throughput (tok/s):          1023.22   
Concurrency:                             1.00      
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   8305.94   
Median E2E Latency (ms):                 8274.94   
P90 E2E Latency (ms):                    8374.91   
P99 E2E Latency (ms):                    8659.60   
---------------Time to First Token----------------
Mean TTFT (ms):                          1119.79   
Median TTFT (ms):                        1113.78   
P99 TTFT (ms):                           1192.66   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          14.40     
Median TPOT (ms):                        14.36     
P99 TPOT (ms):                           15.06     
---------------Inter-Token Latency----------------
Mean ITL (ms):                           14.40     
Median ITL (ms):                         14.32     
P95 ITL (ms):                            14.49     
P99 ITL (ms):                            15.14     
Max ITL (ms):                            484.70    
==================================================

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.