Accelerator PyTorch Code Porting（代码层，名称永不变）

如何特化本 skill（初始化时改写正文，不改名）

下方正文是当前已验证实例（MUSA / MTT S5000，2026-09，面向 CUDA↔MUSA 双平台代码库）。在其他平台首次初始化时按 accel-init 阶段 3 的调研结论改写；改写后本 skill 即该平台版本。已装机器不追仓库新版。

1. API 映射表是本层核心：换成本平台与 CUDA 的真实对照（设备字符串/可见设备变量/通信后端/显存 API/厂商 smi），逐行验证后再写。
2. 生态兼容节（Transformers/加速库：上游原生/厂商 fork/不支持）与扩展编译节按本平台实况重写，附证据链接。
3. 设备探测 helper、AMP、分布式、排障、迁移清单的结构保留，内容换本平台验证过的。
4. 平台名只出现在内容里；本机事实写宿主机层。

This skill guides Python/PyTorch development on the local Moore Threads MTT S5000 platform. It is intended for CUDA-centric repositories that need one maintainable codebase across NVIDIA CUDA and Moore Threads MUSA.

Related global skills have distinct ownership:

• accel-pytorch-python: Python environments, package indexes, and wheel compatibility;
• the host-layer skill (local <host>-environment / system-profile): this server's installed Toolkit paths, driver, hardware, and dynamic-library diagnosis;
• accel-platform-install: portable driver/Toolkit installation on another Ubuntu 22.04 host.

Repository instructions remain authoritative. Read the active repository's AGENTS.md, environment documentation, and experiment rules before editing or running anything. Never use this skill to bypass a repository's restrictions on GPU runs, data preparation, artifact paths, commits, or pushes.

1. Local Platform Facts

Validated on worker31069 on 2026-09-02:

The vendor PyTorch build currently registers torch.musa when torch is imported. Shared cross-platform code should still avoid an unconditional import torch_musa, because that breaks ordinary CUDA/CPU environments.

Optional compatibility/performance packages are not currently installed in the local environment: torchada, Triton, bitsandbytes, flash-attn, and MATE. Do not generate code that depends on them until the target environment has been checked and the user has chosen to install a version matched to this MUSA stack.

2. Core Migration Policy

Use this order of preference:

1. Backend-neutral PyTorch plus a small device abstraction for maintained application code. Prefer .to(device), torch.autocast(device_type=...), and helper functions for synchronization, memory, and seeding.
2. Explicit MUSA implementation behind one platform boundary when behavior or kernels differ. Preserve the CUDA implementation and shared algorithm.
3. torchada for a large legacy CUDA codebase only when it is installed, version-compatible, imported at the entry point before CUDA-dependent modules, and tested end to end. It is an adapter, not proof that every CUDA kernel is correct on MUSA.
4. Native MUSA extension or pure PyTorch fallback for unsupported custom CUDA extensions. Do not silently skip a required op or pretend a CUDA wheel works.

Do not mechanically replace every cuda string. First classify each occurrence as one of:

• device selection;
• visibility/resource ownership;
• AMP or memory API;
• distributed backend;
• CUDA-only package or extension;
• kernel/architecture assumption;
• documentation or artifact metadata;
• a genuine NVIDIA-only path that must remain unchanged.

Keep platform differences in one helper/module whenever possible. Do not fork whole model, quantization, training, or experiment trees for MUSA.

3. Environment Check Before Coding

Run from the intended environment, without loading a model:

mthreads-gmi
musaInfo
mcc --version
python --version
python - <<'PY'
import importlib.metadata
import torch

print("torch:", torch.__version__)
print("torch_musa:", importlib.metadata.version("torch-musa"))
print("torch.version.musa:", getattr(torch.version, "musa", None))
print("musa available:", hasattr(torch, "musa") and torch.musa.is_available())
print("musa devices:", torch.musa.device_count() if hasattr(torch, "musa") else 0)
print("mccl available:", torch.distributed.is_backend_available("mccl"))
PY

If torch.musa.is_available() is false, check the driver, MUSA ToolKit, the matching torch/torch_musa wheels, MUSA_HOME, PATH, and LD_LIBRARY_PATH. Do not debug application code until the runtime is healthy.

Do not install a CUDA-pinned requirements*.txt unchanged on this host. MUSA PyTorch wheels and vendor libraries must match the installed driver/toolkit. Installing an ordinary public PyTorch wheel can replace the MUSA build and pull NVIDIA runtime packages.

4. Device Detection and Backend Abstraction

For new cross-platform code, use a helper such as:

import os

import torch


def accelerator_type() -> str:
    if hasattr(torch, "musa") and torch.musa.is_available():
        return "musa"
    if torch.cuda.is_available():
        return "cuda"
    return "cpu"


def accelerator_device(index: int | None = None) -> torch.device:
    kind = accelerator_type()
    if kind == "cpu":
        return torch.device("cpu")
    if index is None:
        index = int(os.environ.get("LOCAL_RANK", "0"))
    return torch.device(kind, index)


def set_accelerator_device(index: int) -> torch.device:
    device = accelerator_device(index)
    if device.type == "musa":
        torch.musa.set_device(device)
    elif device.type == "cuda":
        torch.cuda.set_device(device)
    return device

Important rules:

• Prefer tensor.to(device) and model.to(device) over .cuda() or .musa().
• A MUSA tensor has tensor.device.type == "musa", never "cuda".
• Replace checks such as device.type == "cuda" with device.type in ("cuda", "musa") only when the branch truly supports both.
• Do not use torch.cuda.is_available() as a generic GPU check. It remains false on MUSA, including when torchada is used.
• Do not select MUSA from a hostname. Detect runtime capability.
• If both backends could be present, use the repository's documented priority; do not change it implicitly.

CUDA-to-MUSA API map

Do not assume CUDA allocator environment variables such as PYTORCH_CUDA_ALLOC_CONF control MUSA memory. Use them only if the installed vendor documentation explicitly says so; otherwise record them as irrelevant to the MUSA process.

5. Minimal Single-Device Program

import torch


def main() -> None:
    if not hasattr(torch, "musa") or not torch.musa.is_available():
        raise RuntimeError("MUSA is unavailable in the active PyTorch runtime")

    device = torch.device("musa", 0)
    torch.musa.set_device(device)

    model = torch.nn.Linear(1024, 1024).to(device=device, dtype=torch.bfloat16)
    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3)
    inputs = torch.randn(8, 1024, device=device, dtype=torch.bfloat16)

    optimizer.zero_grad(set_to_none=True)
    with torch.autocast(device_type="musa", dtype=torch.bfloat16):
        loss = model(inputs).float().square().mean()
    loss.backward()

    grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
    if not torch.isfinite(grad_norm):
        raise FloatingPointError(f"Non-finite gradient norm before Adam: {grad_norm}")
    optimizer.step()
    torch.musa.synchronize()
    print({"device": str(device), "loss": float(loss), "grad_norm": float(grad_norm)})


if __name__ == "__main__":
    main()

This is a platform smoke, not evidence that a real model, fused op, attention backend, optimizer, or multi-hour training run is stable.

6. AMP and Numerical Stability

Use the backend-neutral AMP APIs:

device_type = device.type
with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
    output = model(inputs)
    loss = criterion(output, targets)

For fp16 training when scaling is needed:

scaler = torch.amp.GradScaler(device.type)

with torch.autocast(device_type=device.type, dtype=torch.float16):
    loss = model(inputs).loss
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
if not torch.isfinite(grad_norm):
    raise FloatingPointError("Non-finite gradient norm before optimizer.step()")
scaler.step(optimizer)
scaler.update()

BF16 usually does not need gradient scaling. Keep the scientific precision recipe unchanged during initial porting; change dtype only after isolating a real unsupported or unstable path.

Do not infer backward correctness from a successful forward. For a new model or operator, separately verify:

1. finite forward outputs and loss;
2. finite incoming and accumulated gradients;
3. finite gradient norm before clipping;
4. finite gradients after clipping;
5. finite parameters and optimizer moments after the step;
6. finite and noncollapsed exported state.

Fail before optimizer.step() when incoming gradients or the returned clipping norm are non-finite. A finite scalar loss is insufficient.

Local verified attention limitation

On this exact S5000/MUSA 5.2/PyTorch 2.9.1 stack, Qwen3.5-2B binary-code QAT with Transformers SDPA produced NaN gradients during attention backward. The same P0/data/seed/recipe completed 20 steps with eager attention. For that validated path, select eager before model construction:

config = transformers.AutoConfig.from_pretrained(model_path)
config._attn_implementation = "eager"
model = transformers.AutoModelForCausalLM.from_config(config)

or, when supported by the loader:

model = transformers.AutoModelForCausalLM.from_pretrained(
    model_path,
    attn_implementation="eager",
)

Do not mutate attention mode only after modules are built. Do not generalize this local Qwen3.5 finding into a claim that all MUSA SDPA forward/backward paths are broken; test each model/runtime combination.

7. Transformers and Accelerate

Treat Hugging Face portability as something to verify, not assume.

• Pass an explicit torch.device("musa", index) or move the constructed model with .to(device) for the first single-device port.
• device_map="auto", Accelerate dispatch hooks, quantized loaders, and offload code often contain CUDA-specific branches. Audit them before enabling.
• Do not set dispatch=True merely because a GPU exists; confirm the dispatch implementation supports MUSA.
• Record the requested and resolved attention implementation.
• Use native torch.optim.AdamW / Transformers adamw_torch first. CUDA-only bitsandbytes paged/fused optimizers require a MUSA-compatible build and are not installed in the current environment.
• FlashAttention, xFormers, Triton kernels, BitBLAS, Apex, DeepSpeed fused ops, and vendor-specific quantization packages need explicit MUSA support. Keep a pure PyTorch fallback when practical.
• Validate model load, one forward, one backward, optimizer step, save, and a fresh-process reload as separate gates.

When a config controls implementation choice, set it before from_config()/from_pretrained(). Record what the created submodule actually resolved, not only the CLI request.

8. Multi-GPU with MCCL

This host has validated 1/2/8-rank MCCL all-reduce. A generic DDP entry point:

import os

import torch
import torch.distributed as dist


def main() -> None:
    local_rank = int(os.environ["LOCAL_RANK"])
    torch.musa.set_device(local_rank)
    device = torch.device("musa", local_rank)

    dist.init_process_group(backend="mccl", init_method="env://")
    try:
        model = build_model().to(device)
        model = torch.nn.parallel.DistributedDataParallel(
            model,
            device_ids=[local_rank],
            output_device=local_rank,
        )
        train(model, device)
    finally:
        dist.destroy_process_group()


if __name__ == "__main__":
    main()

Launch from the shell or scheduler-owned wrapper:

MUSA_VISIBLE_DEVICES=0,1 \
torchrun --standalone --nproc-per-node=2 train.py

Rules:

• MUSA_VISIBLE_DEVICES selects physical cards and renumbers them logically. Inside the process, LOCAL_RANK=0 uses logical musa:0.
• Do not map LOCAL_RANK back to a physical ID.
• The scheduler/launcher owns visibility. Worker commands should not overwrite it unless the launcher contract explicitly says they should.
• CUDA_VISIBLE_DEVICES alone does not restrict MUSA devices on this host.
• Use one process per GPU. Avoid DataParallel; prefer DDP.
• Check torch.distributed.is_backend_available("mccl") before launch.
• Ensure every rank reaches collectives in the same order and always destroys the process group after success or failure.
• Use DistributedSampler, call sampler.set_epoch(epoch), and make rank-zero artifact writes explicit.
• FSDP, tensor parallel, pipeline parallel, checkpoint sharding, and elastic recovery need their own S5000 preflight. A successful MCCL all-reduce does not prove those higher layers.

For code intentionally kept CUDA-shaped through torchada, nccl may be mapped to MCCL. For native MUSA code, use mccl explicitly so runtime behavior is auditable.

9. Memory, Timing, and Performance

Use native memory APIs through a backend helper:

def synchronize(device_type: str) -> None:
    if device_type == "musa":
        torch.musa.synchronize()
    elif device_type == "cuda":
        torch.cuda.synchronize()


def max_memory_allocated(device: torch.device) -> int:
    if device.type == "musa":
        return torch.musa.max_memory_allocated(device)
    if device.type == "cuda":
        return torch.cuda.max_memory_allocated(device)
    return 0

• Synchronize before and after timing asynchronous GPU work.
• Record both allocated and reserved peak memory.
• Do not use empty_cache() as an OOM strategy inside the hot loop. Fix tensor lifetime, batch size, checkpointing, or the algorithm.
• Avoid constructing dense intermediate matrices whose size scales as samples x codebook x dimension. Chunk or stream the computation.
• Validate DataLoader pin_memory, nonblocking transfers, multiprocessing, and persistent workers on the actual MUSA runtime before treating them as a performance optimization.
• First establish correctness with native PyTorch ops. Add fused/vendor kernels one at a time with numerical and backward checks.

Performance parity and numerical parity are different. A correct eager fallback can be accepted before a faster MATE/Triton/custom-kernel path is introduced.

10. CUDA Extensions and Third-Party Packages

Native MUSA extension

The torch_musa extension interface mirrors CUDAExtension but uses MUSA tools:

from setuptools import setup
from torch_musa.utils.musa_extension import BuildExtension, MUSAExtension

setup(
    ext_modules=[
        MUSAExtension(
            "my_ops",
            sources=["bindings.cpp", "kernel.mu"],
            extra_compile_args={"cxx": ["-O3"], "mcc": ["-O2"]},
        )
    ],
    cmdclass={"build_ext": BuildExtension},
)

CUDA assumptions that need review include .cu sources, nvcc, CUDA_HOME, sm_XX, CUDA headers, at::cuda, c10::cuda, CUDA streams, cuBLAS/cuDNN/NCCL, and architecture-specific intrinsics.

torch_musa.utils.simple_porting and musa-converter can accelerate source conversion, but generated changes still require review and tests.

torchada

torchada can translate many CUDA Python APIs, NCCL, extensions, and device strings at runtime:

import torch

if getattr(torch.version, "musa", None) is not None:
    import torchada  # Must precede modules that import CUDA extension helpers.

Use torchada when preserving a large CUDA-oriented upstream is more valuable than converting every API. Before choosing it:

1. verify it is installed and supports the exact torch_musa version;
2. inventory unsupported CUDA packages and native libraries;
3. test device checks because torch.cuda.is_available() intentionally remains false and a translated torch.device("cuda") becomes a MUSA device whose .type is "musa";
4. build extensions in a disposable/clean checkout because torchada extension porting may rewrite eligible CUDA sources in place;
5. compare outputs and gradients against a trusted backend.

Do not add torchada to a maintained native-MUSA codebase merely to avoid writing a small device helper.

11. Debugging Backend Failures

Classify the first failure boundary:

1. import/runtime/driver;
2. device allocation or transfer;
3. unsupported operator/dtype/layout;
4. forward numerical divergence;
5. backward numerical divergence;
6. gradient clipping or AMP scaling;
7. optimizer state corruption;
8. distributed collective/hang;
9. save/export/reload mismatch;
10. performance-only regression.

Reduce to the smallest reproducer while preserving the failing dtype, shape, layout, attention implementation, autocast state, and accumulation boundary. Do not change five variables at once.

Useful torch_musa tools:

from torch_musa.utils.compare_tool import CompareWithCPU, NanInfTracker, open_module_tracker

open_module_tracker(model)

with NanInfTracker():
    run_one_step(model)

with CompareWithCPU(atol=1e-3, rtol=1e-3, target_op=["torch.ops.aten.addmm"]):
    run_one_step(model)

CPU comparison can be expensive and may not support every low-precision op. Enable it for a bounded step/rank/op range. For distributed debugging, write separate rank logs and usually enable deep comparison on rank 0 only.

When reporting a MUSA bug, preserve:

• device model and count;
• driver, toolkit, MCCL, Python, torch, torch_musa, and framework versions;
• Git commit and complete command;
• visibility variables and logical rank mapping;
• first failing op/module/tensor;
• shape, stride, dtype, autocast, and attention backend;
• first non-finite phase and index;
• a bounded reproducer and trusted-backend comparison;
• whether eager/pure-PyTorch fallback succeeds.

12. Migration Audit Checklist

Before editing, search the target repository for:

torch.cuda
.cuda(
device="cuda"
device.type == "cuda"
CUDA_VISIBLE_DEVICES
PYTORCH_CUDA_ALLOC_CONF
nccl
nvidia-smi
nvcc
CUDA_HOME
CUDAExtension
bitsandbytes
flash_attn
xformers
triton
BitBLAS
apex
deepspeed

For every match, decide whether to abstract, retain as NVIDIA-only, replace with a MUSA equivalent, add a tested fallback, or mark unsupported.

Before declaring a port ready:

• environment and device discovery pass;
• single-device fp32 and target low-precision tensor ops pass;
• forward, backward, clipping, and optimizer state are finite;
• target attention implementation is explicit and recorded;
• save and fresh reload are verified;
• memory and wall time are measured with synchronization;
• two-rank MCCL passes before scaling to eight ranks;
• all ranks use logical LOCAL_RANK devices exactly once;
• CUDA-only dependencies have explicit handling;
• CUDA behavior remains unchanged in shared code;
• only a bounded preflight is used before user-authorized formal work;
• conclusions distinguish runtime success, numerical parity, and scientific result quality.

13. Agent Working Rules

When this skill is active, the Agent should:

1. inspect the current runtime and repository rules before proposing changes;
2. state whether it is using native MUSA APIs, a shared abstraction, torchada, or a native extension;
3. preserve the existing CUDA path unless the user explicitly removes it;
4. make the smallest compatibility change at the narrowest shared boundary;
5. perform caller/impact analysis before changing shared helpers;
6. verify static syntax first, then follow a staged runtime ladder;
7. never claim an optional library or kernel is available without checking the active environment;
8. never equate successful execution with cross-platform numerical parity;
9. defer expensive/formal runs to the user when repository rules require it;
10. record exact versions, commands, paths, hashes, and limitations in handoff reports.

14. Sources and Freshness

Primary references consulted for this skill:

• torch_musa README: https://github.com/MooreThreads/torch_musa
• torch_musa porting/debug utilities: https://github.com/MooreThreads/torch_musa/blob/main/torch_musa/utils/README.md
• torchada README: https://github.com/MooreThreads/torchada
• Local runtime evidence: the host's MUSA local-stack manifest, if present (see the host-environment skill for its location).

The MUSA ecosystem changes quickly. Re-check upstream documentation and the active environment before relying on package availability, extension support, attention kernels, torch.compile, graph capture, or performance claims.