[Refactor] NPU FLA: delegate to fla native operators, remove MindSpeed/twinkle-own kernels

[ys2025-AI]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

• Remove twinkle-own Triton kernels (causal_conv1d.py, chunk_gated_delta_rule.py) that depended on MindSpeed; delegate to fla's native causal_conv1d and chunk_gated_delta_rule which ship their own Ascend-backend dispatch
• Move npu_causal_conv1d_fn adapter into npu_impls/fla.py; drop redundant weight transpose (fla expects [D,W], not [W,D])
• Simplify builtin.py FLA dependency check: only requires flash-linear-attention (no longer mindspeed>=0.12.1)
• Update gdn_padding_free.py import path for npu_causal_conv1d_fn
• Clean up unused Optional import and fix activation type hint
• Update tests to stub fla package as unavailable

Experiment results

Test environment

Component	Version
Hardware	Ascend NPU A3
CANN	9.0.0
Python	3.11.15
torch	2.9.0+cpu
torch_npu	2.9.0.post2
triton	3.2.0
triton-ascend	3.2.1
transformers	5.9.0
flash-linear-attention	0.5.2

Unit tests

tests/kernel/ + tests/patch/: 66 passed, 5 skipped
test_qwen35_fla_bwd_precision.py (bf16 fla vs fp32 torch): 3 passed

fsdp2 e2e (Qwen3.5-4B, fsdp=8, dp=1, max_length=8192)

step	FLA ON	FLA OFF	diff (ON−OFF)
1	1.5515	1.5520	−0.0005
2	1.3512	1.3517	−0.0005
3	1.6020	1.6018	+0.0002
4	1.6367	1.6358	+0.0009
5	1.6180	1.6192	−0.0012

max |diff| = 0.0012, mean |diff| = 0.0007

sp2 e2e (Qwen3.5-4B, fsdp=8, dp=1, ulysses=2, max_length=8192)

step	FLA ON	FLA OFF	diff (ON−OFF)
0	1.5521	1.5538	−0.0017
1	1.3533	1.3536	−0.0003
2	1.6036	1.6035	+0.0001
3	1.6380	1.6368	+0.0012
4	1.6190	1.6195	−0.0005

max |diff| = 0.0017, mean |diff| = 0.0008

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.