[Feature] Support fusedmoe on Blackwell

[Echo-Nie]

Motivation

Support fusedmoe on Blackwell

Modifications

prerequisite PR: Paddle#76708
custom_ops/gpu_ops/cutlass_kernels/moe_gemm/fused_moe_cutlass_kernel.h
custom_ops/gpu_ops/cutlass_kernels/moe_gemm/fused_moe_gemm_kernels_template.h

Usage or Command

In test_fusedmoe.py, add the following import:

from fastdeploy.model_executor.layers.quantization.weight_only import WeightOnlyConfig, WINT4Config, WINT8Config

Then uncomment the corresponding lines:

quant_config = BlockWiseFP8Config(weight_block_size=[128, 128])
# quant_config = WINT8Config({})
# quant_config = WINT4Config({})

Finally, run:

python FastDeploy/tests/layers/test_fusedmoe.py

Accuracy Tests

BF16

num_token: 4096
[13.9 13.9 13.9 13.9 13.9]
13.9 57.9
0.6 TB/s
num_token: 8192
[27.2 27.2 27.2 27.2 27.2]
27.2 59.2
0.3 TB/s
num_token: 16384
[53.8 53.8 53.8 53.8 53.8]
53.8 59.9
0.1 TB/s
num_token: 32768
[106.9 106.9 106.9 106.9 106.9]
106.9 60.3
0.1 TB/s

BlockWiseFP8Config

num_token: 4096
[9.7 9.7 9.7 9.7 9.7]
9.7 83.0
0.8 TB/s
num_token: 8192
[18.1 18.1 18.1 18.1 18.1]
18.1 89.0
0.4 TB/s
num_token: 16384
[34.8 34.7 34.8 34.7 34.8]
34.8 92.6
0.2 TB/s
num_token: 32768
[68.  68.1 68.1 68.1 68.1]
68.1 94.6
0.1 TB/s

WINT8Config

num_token: 4096
[14.4 14.4 14.4 14.4 14.4]
14.4 55.9
0.6 TB/s
num_token: 8192
[28.3 28.3 28.3 28.3 28.3]
28.3 56.9
0.3 TB/s
num_token: 16384
[55.8 55.9 55.8 55.9 55.8]
55.8 57.7
0.1 TB/s
num_token: 32768
[111.1 111.1 111.1 111.1 111.1]
111.1 58.0
0.1 TB/s

WINT4Config

num_token: 4096
[14.8 14.8 14.8 14.8 14.8]
14.8 54.4
0.5 TB/s
num_token: 8192
[28.9 28.9 28.9 28.9 28.9]
28.9 55.7
0.3 TB/s
num_token: 16384
[57.3 57.3 57.3 57.3 57.2]
57.2 56.3
0.1 TB/s
num_token: 32768
[113.9 113.9 113.9 113.9 113.8]
113.8 56.6
0.1 TB/s

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@532f9ba). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #5325   +/-   ##
==========================================
  Coverage           ?   60.87%           
==========================================
  Files              ?      329           
  Lines              ?    41160           
  Branches           ?     6274           
==========================================
  Hits               ?    25055           
  Misses             ?    14213           
  Partials           ?     1892           

Flag	Coverage Δ
GPU	60.87% <ø> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[paddle-bot]

Thanks for your contribution!

[Echo-Nie]

910 改成 1010，放行 SM100

[Echo-Nie]

91 改成 101，放行 SM100

[zoooo0820]

LGTM