use sgl_per_token_group_quant_fp8 kernel#3493
Merged
Conversation
zhyncs
suggested changes
Feb 11, 2025
5 tasks
zhyncs
approved these changes
Feb 12, 2025
Collaborator
|
@BBuf This kernel only supports Nvidia GPU for now, right? |
Collaborator
Author
Yeah, but I think support on AMD GPUs is also quite simple; it only requires a slight rewrite of this kernel. However, I am not familiar with AMD's development. |
Collaborator
|
Thanks. I think I need to manage some specific cases for AMD GPU when applying this kernel. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
end2end perfomance:
first time
main:
============ Serving Benchmark Result ============ Backend: sglang Traffic request rate: 8.0 Max reqeuest concurrency: not set Successful requests: 1000 Benchmark duration (s): 234.70 Total input tokens: 301701 Total generated tokens: 188375 Total generated tokens (retokenized): 187534 Request throughput (req/s): 4.26 Input token throughput (tok/s): 1285.46 Output token throughput (tok/s): 802.61 Total token throughput (tok/s): 2088.07 Concurrency: 262.34 ----------------End-to-End Latency---------------- Mean E2E Latency (ms): 61571.47 Median E2E Latency (ms): 52159.16 ---------------Time to First Token---------------- Mean TTFT (ms): 2047.59 Median TTFT (ms): 630.04 P99 TTFT (ms): 19879.67 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 580.62 Median TPOT (ms): 461.36 P99 TPOT (ms): 2277.66 ---------------Inter-token Latency---------------- Mean ITL (ms): 319.34 Median ITL (ms): 157.48 P99 ITL (ms): 3867.63 ==================================================pr:
============ Serving Benchmark Result ============ Backend: sglang Traffic request rate: 8.0 Max reqeuest concurrency: not set Successful requests: 1000 Benchmark duration (s): 234.13 Total input tokens: 301701 Total generated tokens: 188375 Total generated tokens (retokenized): 187556 Request throughput (req/s): 4.27 Input token throughput (tok/s): 1288.58 Output token throughput (tok/s): 804.56 Total token throughput (tok/s): 2093.14 Concurrency: 261.91 ----------------End-to-End Latency---------------- Mean E2E Latency (ms): 61322.63 Median E2E Latency (ms): 51756.45 ---------------Time to First Token---------------- Mean TTFT (ms): 2185.78 Median TTFT (ms): 605.41 P99 TTFT (ms): 21702.75 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 593.39 Median TPOT (ms): 484.39 P99 TPOT (ms): 2621.74 ---------------Inter-token Latency---------------- Mean ITL (ms): 318.73 Median ITL (ms): 154.98 P99 ITL (ms): 2892.38 ==================================================second time
main:
============ Serving Benchmark Result ============ Backend: sglang Traffic request rate: 8.0 Max reqeuest concurrency: not set Successful requests: 1000 Benchmark duration (s): 231.16 Total input tokens: 301701 Total generated tokens: 188375 Total generated tokens (retokenized): 187505 Request throughput (req/s): 4.33 Input token throughput (tok/s): 1305.18 Output token throughput (tok/s): 814.92 Total token throughput (tok/s): 2120.10 Concurrency: 245.80 ----------------End-to-End Latency---------------- Mean E2E Latency (ms): 56819.06 Median E2E Latency (ms): 47473.95 ---------------Time to First Token---------------- Mean TTFT (ms): 834.04 Median TTFT (ms): 522.79 P99 TTFT (ms): 4564.58 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 465.60 Median TPOT (ms): 388.09 P99 TPOT (ms): 1439.86 ---------------Inter-token Latency---------------- Mean ITL (ms): 299.96 Median ITL (ms): 157.45 P99 ITL (ms): 2379.64 ==================================================pr:
============ Serving Benchmark Result ============ Backend: sglang Traffic request rate: 8.0 Max reqeuest concurrency: not set Successful requests: 1000 Benchmark duration (s): 229.98 Total input tokens: 301701 Total generated tokens: 188375 Total generated tokens (retokenized): 187525 Request throughput (req/s): 4.35 Input token throughput (tok/s): 1311.87 Output token throughput (tok/s): 819.10 Total token throughput (tok/s): 2130.97 Concurrency: 242.99 ----------------End-to-End Latency---------------- Mean E2E Latency (ms): 55881.33 Median E2E Latency (ms): 46573.34 ---------------Time to First Token---------------- Mean TTFT (ms): 772.20 Median TTFT (ms): 513.58 P99 TTFT (ms): 3324.78 -----Time per Output Token (excl. 1st token)------ Mean TPOT (ms): 447.66 Median TPOT (ms): 388.91 P99 TPOT (ms): 1173.11 ---------------Inter-token Latency---------------- Mean ITL (ms): 295.57 Median ITL (ms): 155.52 P99 ITL (ms): 2161.58 ==================================================Both latency and throughput achieved an end-to-end speedup of approximately 1% in DeepSeek R1.
@zhyncs