[BACKEND] Fix CANN 8.5.1 bishengir compatibility

[Joiin0392]

Description

FlagTree triton_v3.5.x targets LLVM 22 era MLIR syntax, but CANN 8.5.1's bishengir toolchain (2026-02, commit e4e2ba9841d1) is older and cannot parse three MLIR constructs emitted by FlagTree's compiler. This causes all Triton kernel compilations on Ascend to fail, making FlagGems kernels non-functional on CANN 8.5.1.

Root Cause

Three incompatibilities between FlagTree MLIR emission and CANN 8.5.1 bishengir:

1. bufferization.to_tensor result type clause — FlagTree's triton-opt emits bufferization.to_tensor ... to tensor<...> with a result-type clause that bishengir cannot parse.

2. .mlirbc bytecode — FlagTree emits MLIR bytecode (.mlirbc) for inter-tool communication, but bishengir-opt cannot read the newer bytecode version.

3. --link-aicore-bitcode flag — FlagTree passes --link-aicore-bitcode=<path> to bishengir-compile, but CANN 8.5.1's bishengir-compile does not recognize this parameter.

4. Integer pow() dtype — libdevice.pow() passes integer arguments directly to the SIMT path, but bishengir does not support integer pow.

Changes

compiler.py (3 fixes):

1. Add _normalize_mlir_text_for_bishengir(): strips to tensor<...> result-type clause from bufferization.to_tensor lines while keeping restrict writable (required by One-Shot Analysis).

2. Set NPUOptions.use_bytecode = False: use text MLIR instead of bytecode for inter-tool transfer.

3. Remove --link-aicore-bitcode from compile option list (replace with pass).

libdevice.py (1 fix):

Cast integer arguments to fp32 before calling the libdevice SIMT path in pow().

Testing

Environment: Ascend 910B2C x86_64 / CANN 8.5.1 (bishengir 2026-02)

# Before fix: all Triton kernel compilations fail
# After fix: FlagGems kernels compile and execute

python3 -c "
import torch
import flag_gems
flag_gems.enable()
x = torch.randn(1024, device='npu')
y = torch.relu(x)
print('FlagGems kernel OK:', y.shape)
"

Tested end-to-end with vllm-plugin-FL + FlagCX + FlagGems on Qwen3.6-35B-A3B (4-card TP=4):

• FlagGems Triton kernels compile via bishengir ✅
• Inference produces correct output ✅

[CLAassistant]

All committers have signed the CLA.