[Feature][OP] Add V100 (SM70) GPU Support

[mattheliu]

Motivation

为 FastDeploy 添加 NVIDIA V100 GPU (SM70 架构) 支持，使其能在旧版 GPU 上进行开发测试。由于 V100 不支持以下特性，需要同时适配编译系统和运行时逻辑：

• BF16 数据类型：需要 SM80+ (Ampere)
• FP8 量化：需要 SM89+ (Ada Lovelace)
• cp.async 指令：需要 SM80+ (Ampere)，影响 Append Attention 和 MLA Attention
• Marlin GEMM：需要 SM80+ (Ampere)

Modifications

编译系统

• setup_ops.py: 支持 SM70+ 编译，分离 SM70/SM80+ 特有代码
• cpp_extensions.cc: 添加 ENABLE_APPEND_ATTENTION 和 ENABLE_BF16 宏控制条件编译

CUDA Kernel

• gelu_tanh.cu: 修复 tanh.approx.f32 PTX 指令在 SM70 的编译问题
• moe_wna16_marlin_*.cu/h: 修复 Marlin GEMM 模板在 SM70 的编译兼容性

Python 运行时层

• fastdeploy/platforms/cuda.py:

  • 添加 SM 版本检测方法 (get_sm_version())
  • 添加硬件能力检查 (supports_bf16(), supports_fp8(), supports_async_copy(), supports_marlin())
  • Attention backend 自动 fallback (APPEND_ATTN/MLA_ATTN → FLASH_ATTN)

• fastdeploy/config.py: BF16→FP16 dtype 自动降级

• fastdeploy/model_executor/layers/moe/moe.py:

  • Marlin MoE backend → CUTLASS fallback (SM<80)
  • Triton MoE backend → CUTLASS fallback (SM<80)

• fastdeploy/model_executor/layers/quantization/__init__.py:

  • FP8 量化方法自动 fallback (block_wise_fp8→wint8, w4afp8→wint4)

• fastdeploy/model_executor/layers/quantization/mix_quant.py:

  • MixQuantConfig 中 FP8 quant type 自动 fallback

• fastdeploy/model_executor/layers/quantization/weight_only.py:

  • WeightOnlyConfig 中 Marlin/Triton backend fallback

• fastdeploy/model_executor/layers/quantization/block_wise_fp8.py:

  • deep_gemm 导入保护 (SM<89 时跳过)

• attention/ops/*.py: 为 SM80+ 专属 ops 添加 try-except 保护

测试

• tests/layers/test_attention_layer.py: 添加 FP8 SM89+ skip 装饰器
• tests/layers/test_fusedmoe.py: 添加 FP8 SM89+ skip 装饰器
• tests/quantization/test_w4afp8.py: 添加 FP8 SM89+ skip 装饰器
• tests/layers/test_ffn.py: 根据 SM 版本自动选择 dtype 和量化配置

SM70 Fallback 策略总览

功能	原始	SM70 Fallback	原因
数据类型	BF16	FP16	BF16 需要 SM80+
Attention Backend	APPEND_ATTN	FLASH_ATTN	cp.async 需要 SM80+
Attention Backend	MLA_ATTN	FLASH_ATTN	cp.async 需要 SM80+
MoE Backend	Marlin	CUTLASS	Marlin 需要 SM80+
MoE Backend	Triton	CUTLASS	tritonmoe_preprocess 需要 SM80+
量化	block_wise_fp8	wint8	FP8 需要 SM89+
量化	w4afp8	wint4	FP8 需要 SM89+
量化	wfp8afp8	wint8	FP8 需要 SM89+
量化	tensor_wise_fp8	wint8	FP8 需要 SM89+

Usage or Command

# 编译 (指定 SM70 架构)
cd custom_ops && bash build.sh 1 python false [70]

# 或使用 setup_ops.py
cd custom_ops && python setup_ops.py install

# 运行测试
bash test_sm70_compat.sh
# 或单独运行
pytest tests/platforms/test_platforms.py -v
pytest tests/layers/test_attention_layer.py -v
pytest tests/layers/test_ffn.py -v
pytest tests/layers/test_fusedmoe.py -v
pytest tests/quantization/test_w4afp8.py -v
pytest tests/quantization/ -v

Accuracy Tests

V100 (SM70) 上测试结果：

=== 1. Platform Detection ===
current sm_version=70
Platform: CUDAPlatform
Is V100 (SM70): True

=== 2. Platform Tests ===
29 passed, 0 failed

=== 3. Attention Tests ===
5 passed, 1 skipped (FP8 quantization requires SM89+)

=== 4. FFN Tests ===
1 passed, 0 failed

=== 5. MoE Tests ===
1 skipped (FP8 quantization requires SM89+)

=== 6. W4AFP8 Quantization Tests ===
6 passed, 5 skipped (FP8 ops require SM89+)

=== 7. All Quantization Tests ===
46 passed, 9 skipped

=== 8. Non-FP8 Quantization Tests ===
36 passed, 1 skipped (XPU)

Total: 123+ passed, 17 skipped, 0 failed

所有 FP8 相关测试在 V100 上正确跳过（显示 SKIPPED (FP8 ops require SM89+)），非 FP8 功能全部通过。

Checklist

• Add at least a tag in the PR title.
• 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.

[paddle-bot]

Thanks for your contribution!

[mattheliu]

/Re-run failed jobs

[codecov-commenter]

Codecov Report

❌ Patch coverage is 43.21608% with 113 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@9b0a82c). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...loy/model_executor/layers/quantization/__init__.py	28.57%	23 Missing and 2 partials ⚠️
...oy/model_executor/layers/quantization/mix_quant.py	39.28%	15 Missing and 2 partials ⚠️
fastdeploy/platforms/cuda.py	68.29%	11 Missing and 2 partials ⚠️
.../model_executor/layers/quantization/weight_only.py	21.42%	7 Missing and 4 partials ⚠️
..._executor/layers/moe/fused_moe_deepgemm_backend.py	11.11%	8 Missing ⚠️
...del_executor/layers/quantization/block_wise_fp8.py	42.85%	6 Missing and 2 partials ⚠️
fastdeploy/model_executor/layers/moe/moe.py	22.22%	4 Missing and 3 partials ⚠️
...l_executor/layers/moe/fused_moe_cutlass_backend.py	40.00%	6 Missing ⚠️
fastdeploy/config.py	62.50%	1 Missing and 2 partials ⚠️
...cutor/layers/attention/ops/flash_mask_attention.py	0.00%	3 Missing ⚠️
... and 5 more

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6306   +/-   ##
==========================================
  Coverage           ?   67.53%           
==========================================
  Files              ?      389           
  Lines              ?    52098           
  Branches           ?     8116           
==========================================
  Hits               ?    35184           
  Misses             ?    14344           
  Partials           ?     2570           

Flag	Coverage Δ
GPU	67.53% <43.21%> (?)

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.