[Ascend] add vendor:ascend backend

[Sans1J]

Description

Add vendor:ascend backend. The latest code has been synchronized and pre-commit has been executed locally.

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[Sans1J]

Here is the TE-FL related configuration information for running Qwen3-0.6B. Please check if it meets your expectations.

FlagOS-Related Arguments for unified computation backend and communication backend

te_fl_prefer: vendor #flagos # enable flagos:triton in transformer engine fl
te_fl_per_op: "rmsnorm_fwd=vendor:ascend|flagos" # enable custom ops selection for transformer engine fl
te_fl_allow_vendors: "ascend" # allow vendors for transformer engine fl
te_fl_deny_vendors: "nvidia" # deny vendors for transformer engine fl
enable_flag_gems: False #True # enable flag gems to replace torch ops for distributed training
flag_gems_log_path: xxx/TE-FL-test/FlagScale/log_gems/gems.log # path of flag gems logging
flag_gems_unused: [to, copy] # flag gems unused ops list
distributed_backend: nccl #flagcx # enable flagcx for distributed training

Here are the TE-FL-related information from the logs printed during model training.

[default0]:training ...
[default0]:Setting rerun_state_machine.current_iteration to 0...
[default0]:[before the start of training step] datetime: 2026-02-27 09:26:42
[default0]:ninja: no work to do.
[default0]:[2026-02-27 09:27:15,708 TE-FL manager.py:417 INFO] Op 'multi_tensor_adam' using 'default.flagos' (kind=flagos, vendor=None)
[default0]:[WARNING] Please DO NOT tune args ['num_warps']!
[default0]:[WARNING] Please DO NOT tune args ['num_warps']!

[lxd-cumt]

Here is the TE-FL related configuration information for running Qwen3-0.6B. Please check if it meets your expectations.

FlagOS-Related Arguments for unified computation backend and communication backend

te_fl_prefer: vendor #flagos # enable flagos:triton in transformer engine fl te_fl_per_op: "rmsnorm_fwd=vendor:ascend|flagos" # enable custom ops selection for transformer engine fl te_fl_allow_vendors: "ascend" # allow vendors for transformer engine fl te_fl_deny_vendors: "nvidia" # deny vendors for transformer engine fl enable_flag_gems: False #True # enable flag gems to replace torch ops for distributed training flag_gems_log_path: xxx/TE-FL-test/FlagScale/log_gems/gems.log # path of flag gems logging flag_gems_unused: [to, copy] # flag gems unused ops list distributed_backend: nccl #flagcx # enable flagcx for distributed training

Here are the TE-FL-related information from the logs printed during model training.

[default0]:training ... [default0]:Setting rerun_state_machine.current_iteration to 0... [default0]:[before the start of training step] datetime: 2026-02-27 09:26:42 [default0]:ninja: no work to do. [default0]:[2026-02-27 09:27:15,708 TE-FL manager.py:417 INFO] Op 'multi_tensor_adam' using 'default.flagos' (kind=flagos, vendor=None) [default0]:[WARNING] Please DO NOT tune args ['num_warps']! [default0]:[WARNING] Please DO NOT tune args ['num_warps']!

• Could you provide a complete training log so we can verify whether other TE-related ops were executed, such as generic_temm?
• Why is te_fl_prefer: vendor configured instead of flagos?
• If te_fl_prefer: vendor, why did multi_tensor_adam execute the FlagOS implementation?
• Why is enable_flag_gems: false? This will prevent some torch ops from being replaced with FlagGems.

[lxd-cumt]

Please config: transformer_impl: transformer_engine to use TransformerEngine-FL, and furthermore, adapt multiple FlagOS OPs in TE-FL.
Recommend config as follows:

transformer_impl: transformer_engine
te_fl_prefer: flagos
enable_flag_gems: true
flag_gems_log_path: xxx