Skip to content

perf: flydsl gdr decode default on#752

Open
xytpai wants to merge 5 commits into
mainfrom
xyt/flydsl_gdr_on
Open

perf: flydsl gdr decode default on#752
xytpai wants to merge 5 commits into
mainfrom
xyt/flydsl_gdr_on

Conversation

@xytpai
Copy link
Copy Markdown
Contributor

@xytpai xytpai commented May 11, 2026

Qwen3-Next-80B-A3B-Instruct-FP8 TP1 CC64

before

============ Serving Benchmark Result ============
Successful requests:                     640       
Failed requests:                         0         
Maximum request concurrency:             64        
Benchmark duration (s):                  214.48    
Total input tokens:                      5291983   
Total generated tokens:                  641529    
Request throughput (req/s):              2.98      
Output token throughput (tok/s):         2991.10   
Peak output token throughput (tok/s):    5184.00   
Peak concurrent requests:                71.00     
Total token throughput (tok/s):          27664.73  
---------------Time to First Token----------------
Mean TTFT (ms):                          645.40    
Median TTFT (ms):                        296.69    
P99 TTFT (ms):                           6822.79   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          19.96     
Median TPOT (ms):                        20.23     
P99 TPOT (ms):                           26.63     
---------------Inter-token Latency----------------
Mean ITL (ms):                           19.84     
Median ITL (ms):                         12.48     
P99 ITL (ms):                            181.58    
----------------End-to-end Latency----------------
Mean E2EL (ms):                          20505.57  
Median E2EL (ms):                        20441.54  
P99 E2EL (ms):                           37889.31  
==================================================

after

============ Serving Benchmark Result ============
Successful requests:                     640       
Failed requests:                         0         
Maximum request concurrency:             64        
Benchmark duration (s):                  210.18    
Total input tokens:                      5291983   
Total generated tokens:                  641529    
Request throughput (req/s):              3.04      
Output token throughput (tok/s):         3052.27   
Peak output token throughput (tok/s):    5244.00   
Peak concurrent requests:                71.00     
Total token throughput (tok/s):          28230.46  
---------------Time to First Token----------------
Mean TTFT (ms):                          661.36    
Median TTFT (ms):                        294.48    
P99 TTFT (ms):                           6615.41   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          19.53     
Median TPOT (ms):                        19.74     
P99 TPOT (ms):                           24.88     
---------------Inter-token Latency----------------
Mean ITL (ms):                           19.45     
Median ITL (ms):                         12.22     
P99 ITL (ms):                            178.31    
----------------End-to-end Latency----------------
Mean E2EL (ms):                          20098.61  
Median E2EL (ms):                        20026.14  
P99 E2EL (ms):                           37432.80  
==================================================

Qwen3-Next-80B-A3B-Instruct-FP8 TP4 CC64

before

============ Serving Benchmark Result ============
Successful requests:                     640       
Failed requests:                         0         
Maximum request concurrency:             64        
Benchmark duration (s):                  136.24    
Total input tokens:                      5291983   
Total generated tokens:                  641529    
Request throughput (req/s):              4.70      
Output token throughput (tok/s):         4708.83   
Peak output token throughput (tok/s):    7872.00   
Peak concurrent requests:                74.00     
Total token throughput (tok/s):          43552.01  
---------------Time to First Token----------------
Mean TTFT (ms):                          406.49    
Median TTFT (ms):                        195.60    
P99 TTFT (ms):                           3715.26   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          12.50     
Median TPOT (ms):                        12.68     
P99 TPOT (ms):                           15.28     
---------------Inter-token Latency----------------
Mean ITL (ms):                           12.46     
Median ITL (ms):                         8.27      
P99 ITL (ms):                            122.44    
----------------End-to-end Latency----------------
Mean E2EL (ms):                          12867.63  
Median E2EL (ms):                        12935.14  
P99 E2EL (ms):                           23693.25  
==================================================

after

============ Serving Benchmark Result ============
Successful requests:                     640       
Failed requests:                         0         
Maximum request concurrency:             64        
Benchmark duration (s):                  135.11    
Total input tokens:                      5291983   
Total generated tokens:                  641529    
Request throughput (req/s):              4.74      
Output token throughput (tok/s):         4748.10   
Peak output token throughput (tok/s):    7872.00   
Peak concurrent requests:                73.00     
Total token throughput (tok/s):          43915.23  
---------------Time to First Token----------------
Mean TTFT (ms):                          406.24    
Median TTFT (ms):                        196.19    
P99 TTFT (ms):                           3634.30   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          12.41     
Median TPOT (ms):                        12.58     
P99 TPOT (ms):                           15.43     
---------------Inter-token Latency----------------
Mean ITL (ms):                           12.37     
Median ITL (ms):                         8.20      
P99 ITL (ms):                            122.39    
----------------End-to-end Latency----------------
Mean E2EL (ms):                          12761.78  
Median E2EL (ms):                        12802.25  
P99 E2EL (ms):                           23315.74  
==================================================

@xytpai xytpai changed the title flydsl gdr decode default on perf: flydsl gdr decode default on May 11, 2026
@xytpai xytpai requested a review from ganyi1996ppo May 12, 2026 04:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants