fla编译triton算子错误

你好，我在使用c500时测试了几个fla版本(都是沐曦的版本)都遇到了如图所示的相关错误
mx-smi回显也如图，错误的具体traceback如下：
[rank0]: Traceback (most recent call last):

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/openvla/vla-scripts/distill_train_stage2.py", line 1453, in <module>

[rank0]: distill_train()

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/draccus/argparsing.py", line 228, in wrapper_inner

[rank0]: response = fn(cfg, args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/openvla/vla-scripts/distill_train_stage2.py", line 1235, in distill_train

[rank0]: student_output = student_vla(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl

[rank0]: return self._call_impl(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl

[rank0]: return forward_call(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 20, in wrapped_fn

[rank0]: ret_val = func(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2030, in forward

[rank0]: loss = self.module(inputs, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl

[rank0]: return self._call_impl(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1845, in _call_impl

[rank0]: return inner()

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1793, in inner

[rank0]: result = forward_call(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/openvla/prismatic/extern/hf/modeling_prismatic.py", line 404, in forward

[rank0]: language_model_output = self.language_model(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl

[rank0]: return self._call_impl(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl

[rank0]: return forward_call(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.py", line 1189, in forward

[rank0]: outputs = self.model(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl

[rank0]: return self._call_impl(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl

[rank0]: return forward_call(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.py", line 988, in forward

[rank0]: layer_outputs = self._gradient_checkpointing_func(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/_compile.py", line 32, in inner

[rank0]: return disable_fn(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 745, in _fn

[rank0]: return fn(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/utils/checkpoint.py", line 496, in checkpoint

[rank0]: ret = function(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl

[rank0]: return self._call_impl(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl

[rank0]: return forward_call(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/openvla/prismatic/models/backbones/llm/llama_gated_delta.py", line 220, in forward

[rank0]: attn_output, present_key_value = self.self_attn(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl

[rank0]: return self._call_impl(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl

[rank0]: return forward_call(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/openvla/prismatic/models/backbones/llm/gated_delta/gated_delta_net.py", line 289, in forward

[rank0]: o, new_recurrent_state = chunk_gated_delta_rule(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 745, in _fn

[rank0]: return fn(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/fla/ops/gated_delta_rule/chunk.py", line 313, in chunk_gated_delta_rule

[rank0]: o, final_state = ChunkGatedDeltaRuleFunction.apply(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/autograd/function.py", line 575, in apply

[rank0]: return super().apply(args, *kwargs) # type: ignore[misc]

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/fla/utils.py", line 164, in wrapper

[rank0]: return fn(contiguous_args, *contiguous_kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/torch/amp/autocast_mode.py", line 503, in decorate_fwd

[rank0]: return fwd(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/fla/ops/gated_delta_rule/chunk.py", line 174, in forward

[rank0]: g, o, A, final_state = chunk_gated_delta_rule_fwd(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/fla/ops/gated_delta_rule/chunk.py", line 31, in chunk_gated_delta_rule_fwd

[rank0]: A = chunk_scaled_dot_kkt_fwd(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/fla/ops/common/chunk_scaled_dot_kkt.py", line 114, in chunk_scaled_dot_kkt_fwd

[rank0]: chunk_scaled_dot_kkt_fwd_kernel[(NT, B * H)](

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/jit.py", line 345, in <lambda>

[rank0]: return lambda args, kwargs: self.run(grid=grid, warmup=False, args, **kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/autotuner.py", line 396, in run

[rank0]: return self.fn.run(args, *kwargs)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/autotuner.py", line 212, in run

[rank0]: timings = {config: self._bench(args, config=config, *kwargs) for config in pruned_configs}

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/autotuner.py", line 212, in <dictcomp>

[rank0]: timings = {config: self._bench(args, config=config, *kwargs) for config in pruned_configs}

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/autotuner.py", line 137, in _bench

[rank0]: return do_bench(kernel_call, warmup=self.num_warmups, rep=self.num_reps, quantiles=(0.5, 0.2, 0.8))

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/testing.py", line 152, in do_bench

[rank0]: fn()

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/autotuner.py", line 118, in kernel_call

[rank0]: self.fn.run(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/runtime/jit.py", line 662, in run

[rank0]: kernel = self.compile(

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/compiler/compiler.py", line 283, in compile

[rank0]: next_module = compile_ir(module, metadata)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/backends/metax/compiler.py", line 419, in <lambda>

[rank0]: stages["mlir"] = lambda src, metadata: self.make_mlir(src, metadata, options, self.capability)

[rank0]: File "/mnt/afs/lixiaoou/intern/wanyang/envs/openvla-triton-test-fla040/lib/python3.10/site-packages/triton/backends/metax/compiler.py", line 328, in make_mlir

[rank0]: pm.run(mod)

[rank0]: RuntimeError: PassManager::run failed

好像是对于一些高级的指令比如异步内存拷贝指令不支持

尊敬的开发者您好，请参考【发帖前必看！！！】镜像使用问题提问模版 | 公共 | 沐曦开发者论坛 developer.metax-tech.com/forum/t/fa-tie-qian-bi-kan-jing-xiang-shi-yong-wen-ti-ti-wen-mo-ban/267/ 描述详细使用过程以及日志

一、软硬件信息
1. 服务器厂家：IEIT SYSTEMS
2. 沐曦 GPU 型号：MetaX C500，共 8 卡，单卡显存 65536 MiB
3. 操作系统内核版本：5.14.0-284.25.1.el9_2.x86_64
4. 是否开启 CPU 虚拟化：lscpu 显示 VT-x；但容器内无法确认宿主机是否运行在虚拟化环境中
5. mx-smi 回显关键信息：
- mx-smi version: 2.2.8
- Kernel Mode Driver Version: 3.0.11
- MACA Version: 3.5.3.18
- BIOS Version: 1.27.5.0
- Attached GPUs: 8
- GPU Name: MetaX C500
6. docker info 回显：容器内无法访问 Docker daemon，报错为

Cannot connect to the Docker daemon at unix:///var/run/docker.sock. Is the docker daemon running?

1. 镜像版本：容器内无法直接获取，需宿主机侧补充
2. 启动容器命令：容器内无法获取原始 docker run / K8s 启动命令，需宿主机侧补充
3. 容器内执行命令：

**二、问题现象**
1. 问题描述：在沐曦 `MetaX C500` 上运行 OpenVLA 的 GatedDelta/FLA 路径时，部分 Triton kernel 在编译阶段失败，错误发生在 `fla` 的 GatedDelta chunk kernel 上，不是数据错误、OOM、NCCL 通信错误，也不是 teacher 模型损坏。
3. 关键报错日志如下：
```text
loc(".../fla/ops/common/chunk_scaled_dot_kkt.py":62:22): error:
failed to legalize operation 'triton_gpu.async_commit_group'
that was explicitly marked illegal

RuntimeError: PassManager::run failed

1. 现象判断：
2. 报错点在 Triton/MLIR lowering 阶段
3. 失败指令是 triton_gpu.async_commit_group
4. 这说明问题更接近“MetaX Triton backend 对该 kernel 的 async/N-buffer lowering 不支持或不稳定”，而不是 Python 业务逻辑错误

三、希望专家协助确认的问题
1. MACA 3.5.3.18 / KMD 3.0.11 是否官方支持 triton_gpu.async_commit_group 在 MetaX Triton backend 上的 lowering？
3. 对 fla 的以下 kernel：
- chunk_scaled_dot_kkt.py
- chunk_delta_h.py
- chunk_o.py
- wy_fast.py
是否有推荐的 MetaX 专用编译参数、补丁，或已知不支持项？
4. 如果 async_commit_group 是当前后端已知不支持的指令，是否有官方建议的规避方式，例如：
- 禁用 async pipeline
- 使用保守的 num_warps/num_stages
- 升级到指定的 torch + triton + MACA 版本组合

尊敬的开发者您好，请给出您在容器内运行代码的详细命令

命令在文件中

您看看还需要补充什么信息吗

尊敬的开发者您好，请使用沐曦开发者社区软件下载中心适配的trition安装包尝试，async_commit_group正在适配中，等待后续版本支持