[CI]【Hackathon 10th Spring No.45-part2】Add SM75/SM80 compile guards for cutlass and MoE tail ops

[bobby-cloudforge]

Motivation

PR #6488 added conditional compilation guards to cpp_extensions.cc for most SM80+ and SM75+ ops, enabling FastDeploy to compile and run on T4 (SM75) and V100 (SM70) hardware. However, 12 ops added in later commits (post-PR #6488 branch point) were not covered by those guards:

• 5 SM75+ ops (cutlass_scaled_mm, cutlass_scaled_mm_azp, static_scaled_fp8_quant, dynamic_scaled_fp8_quant, dynamic_per_token_scaled_fp8_quant) — only compiled when ENABLE_SCALED_MM_C2X / ENABLE_SM75_EXT_OPS is set (SM75+), yet registered unconditionally in the module
• 7 SM80+ ops (prefill_permute_to_masked_gemm, depermute_prefill_combine, radix_topk_ragged_transform, dsk_attn_write_cache, indexer_k_quant_and_cache, cp_gather_indexer_k_quant_cache, per_token_group_fp8_quant) — from DSA/MoE kernel sources only compiled at SM80+, but missing guards in the pybind module

Without these guards, python setup.py build_ext fails with undefined symbol link errors on T4/V100 targets.

Modifications

Single file changed: custom_ops/gpu_ops/cpp_extensions.cc (+8 lines, -13 lines net)

Block 1 — SM75+ ops (wraps 5 ops with #ifdef ENABLE_SM75_EXT_OPS / #endif):

Op	Source
cutlass_scaled_mm	w8a8/scaled_mm_entry.cu
cutlass_scaled_mm_azp	w8a8/scaled_mm_entry.cu
static_scaled_fp8_quant	quantization/common.cu
dynamic_scaled_fp8_quant	quantization/common.cu
dynamic_per_token_scaled_fp8_quant	quantization/common.cu

Block 2 — SM80+ tail ops (wraps 7 ops with #ifdef ENABLE_SM80_EXT_OPS / #endif):

Op	Source
prefill_permute_to_masked_gemm	moe/prefill_permute_to_masked_gemm.cu
depermute_prefill_combine	moe/depermute_prefill_combine.cu
radix_topk_ragged_transform	sparse_indexer/indexer_topk.cu
dsk_attn_write_cache	append_attn/ds_mla_cache_kernel.cu
indexer_k_quant_and_cache	append_attn/ds_mla_cache_kernel.cu
cp_gather_indexer_k_quant_cache	append_attn/ds_mla_cache_kernel.cu
per_token_group_fp8_quant	append_attn/ds_mla_cache_kernel.cu

Guard balance after this PR: #if*=18, #endif=18 — balanced.

Usage or Command

# Build for V100 (SM70) — SM75 and SM80 op groups excluded
CUDA_VISIBLE_DEVICES=0 python setup.py build_ext --inplace

# Build for T4 (SM75) — SM80 op group excluded, SM75 quant ops included
CUDA_VISIBLE_DEVICES=0 python setup.py build_ext --inplace

# Build for A100+ (SM80+) — all op groups included
CUDA_VISIBLE_DEVICES=0 python setup.py build_ext --inplace

Guard activation is driven by the compile macros already defined in setup_ops.py by PR #6488 (ENABLE_SM75_EXT_OPS at cc≥75, ENABLE_SM80_EXT_OPS at cc≥80). No changes to setup_ops.py are needed in this PR.

Accuracy Tests

This PR only modifies conditional compilation guards in the pybind11 module registration block. No kernel logic, numerical computation, or runtime dispatch is altered. On SM80+ hardware (where all macros are defined), the full op set remains registered and behavior is identical to before this PR.

Checklist

• No functional logic is changed — only #ifdef/#endif guard insertion
• Uses existing macros defined by PR 【Hackathon 10th Spring No.45】FastDeploy 支持在 T4/V100 硬件的编译 -part #6488 (ENABLE_SM75_EXT_OPS, ENABLE_SM80_EXT_OPS)
• Guard block count balanced (#if*=18, #endif=18)
• Complement to PR 【Hackathon 10th Spring No.45】FastDeploy 支持在 T4/V100 硬件的编译 -part #6488 — covers the 12 ops that PR left unguarded
• CI passes on SM89/SM90 hardware (all guards defined → full op set)

[paddle-bot]

Thanks for your contribution!

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-20 14:03 CST

📋 Review 摘要

PR 概述：为 cpp_extensions.cc 中 12 个遗漏的 pybind11 op 注册添加 SM75+/SM80+ 条件编译 guard，修复在 T4/V100 等低算力 GPU 上的 undefined symbol 链接错误。
变更范围：custom_ops/gpu_ops/cpp_extensions.cc（+8 行，-13 行净变更）
影响面 Tag：OP CI

问题

未发现阻塞性问题。

✅ 验证项

检查点	结果
Guard 宏与已有模式一致（ENABLE_SM75_EXT_OPS / ENABLE_SM80_EXT_OPS）	✅ 与 PR #6488 定义的宏匹配
#if* / #endif 配对平衡	✅ 18/18 平衡
Python 调用层兼容性	✅ 所有被 guard 的 ops 在 Python 层均通过动态 import 引入，低 SM 硬件上不会触发未定义符号
5 个 SM75+ ops 覆盖完整	✅ cutlass_scaled_mm、cutlass_scaled_mm_azp、static_scaled_fp8_quant、dynamic_scaled_fp8_quant、dynamic_per_token_scaled_fp8_quant
7 个 SM80+ ops 覆盖完整	✅ prefill_permute_to_masked_gemm、depermute_prefill_combine、radix_topk_ragged_transform、dsk_attn_write_cache、indexer_k_quant_and_cache、cp_gather_indexer_k_quant_cache、per_token_group_fp8_quant

总体评价

改动范围清晰、逻辑正确，是对 PR #6488 的有效补充。仅涉及 pybind11 模块注册块的条件编译 guard 插入，不改变任何 kernel 逻辑或运行时行为，SM80+ 硬件上行为与变更前完全一致。