[perf] Fun-ASR-Nano fine-tuning on B200: 0.244 s to 0.068 s per step at 1 GPU, mostly from cuDNN SDPA plan builds on every new batch shape (#3705)

* fun_asr_nano: opt-in scoped SDPA backends and compiled decoder for fine-tuning

* tests: FunASRNano opt-ins leave defaults untouched

* fun_asr_nano: declare the torch 2.3 floor the fine-tuning recipe needs

Author: TarzanZhao

examples/industrial_data_pretraining/fun_asr_nano/README.md

@@ -58,6 +58,8 @@ cd Fun-ASR
 pip install -r requirements.txt
 ```
 
+`finetune.sh` sets `llm_conf.sdpa_backends`, which needs `torch.nn.attention`, so fine-tuning with it requires PyTorch 2.3 or newer.
+
 <a name="usage-tutorial"></a>
 
 # TODO

examples/industrial_data_pretraining/fun_asr_nano/README_zh.md

@@ -58,6 +58,8 @@ cd Fun-ASR
 pip install -r requirements.txt
 ```
 
+`finetune.sh` 会设置 `llm_conf.sdpa_backends`，该选项依赖 `torch.nn.attention`，因此使用它进行微调需要 PyTorch 2.3 或更高版本。
+
 <a name="用法教程"></a>
 
 # TODO

examples/industrial_data_pretraining/fun_asr_nano/docs/finetune.md

@@ -87,6 +87,13 @@ For more detailed parameters, refer to: [SenseVoice Model Training and Testing](
 bash finetune.sh
 ```
 
+### Training-speed options
+
+Two `llm_conf` keys, both off by default; `finetune.sh` turns them on.
+
+- `++llm_conf.torch_compile=true` runs the LLM decoder stack through `torch.compile(dynamic=True)` for inputs on a CUDA GPU (a CPU decoder and single-sequence batches stay eager); the first compiled step pays a one-time compile.
+- `++llm_conf.sdpa_backends=[flash,efficient,math]` runs the LLM forward under `torch.nn.attention.sdpa_kernel` with these attention backends. It matters on sm_90 / sm_100 GPUs, where torch prefers cuDNN and cuDNN builds an execution plan for every new batch shape. The flags it sets are process-wide while the forward runs (restored on return), so a thread running attention concurrently in the same process sees the same selection; leave it unset when several models share one process.
+
 ### Recommended Configuration
 
 - For training data less than 1000 hours, it is recommended to fine-tune the audio_adaptor.

examples/industrial_data_pretraining/fun_asr_nano/finetune.sh

@@ -62,4 +62,6 @@ ${train_tool} \
 ++audio_encoder_conf.freeze=true \
 ++audio_adaptor_conf.freeze=true \
 ++llm_conf.freeze=false \
+++llm_conf.torch_compile=true \
+++llm_conf.sdpa_backends=[flash,efficient,math] \
 ++output_dir="${output_dir}" &> ${log_file}

examples/industrial_data_pretraining/fun_asr_nano/model.py

@@ -20,6 +20,7 @@
     normalize_checkpoint_state,
 )
 from funasr.models.fun_asr_nano.device_utils import resolve_autocast_device_type
+from funasr.models.fun_asr_nano.llm_forward_opts import configure_llm_forward
 
 
 from ctc import CTC
@@ -93,6 +94,11 @@ def __init__(
 
         self.llm_dtype = llm_conf.get("llm_dtype", "fp32")
         self.llm = model.to(dtype_map[self.llm_dtype])
+        # Opt-in training-speed switches, both off by default (finetune.sh turns them on):
+        # llm_conf.sdpa_backends sets the process-wide SDPA backend flags while this forward
+        # runs and restores them on return; llm_conf.torch_compile compiles the decoder stack
+        # for inputs on a CUDA device (a CPU decoder stays eager). See llm_forward_opts.py.
+        configure_llm_forward(self.llm, llm_conf)
         llm_dim = model.get_input_embeddings().weight.shape[-1]
 
         # adaptor

examples/industrial_data_pretraining/fun_asr_nano/requirements.txt

@@ -1,4 +1,4 @@
-torch>=2.0
+torch>=2.3
 funasr
 websockets>=12.0
 regex

funasr/models/fun_asr_nano/llm_forward_opts.py

@@ -0,0 +1,66 @@
+"""Opt-in switches for the Fun-ASR-Nano LLM forward, read from ``llm_conf``; both default to off.
+
+``llm_conf.sdpa_backends``: a list of ``flash`` / ``efficient`` / ``math``. The decoder forward runs
+under ``torch.nn.attention.sdpa_kernel`` with these backends: the process-wide SDPA flags are set on
+entry and restored on return, so another thread calling scaled_dot_product_attention meanwhile sees
+the same selection (leave it unset when several models share one process concurrently).
+``llm_conf.torch_compile``: run the decoder stack through ``torch.compile(dynamic=True)`` for
+inputs on a CUDA device, decided per call; a CPU decoder and a 1-sequence batch stay eager.
+"""
+
+import torch
+
+SDPA_BACKENDS = {"flash": "FLASH_ATTENTION", "efficient": "EFFICIENT_ATTENTION", "math": "MATH"}
+
+
+def resolve_sdpa_backends(names):
+    """``llm_conf.sdpa_backends`` (a list of names, or ``None``) -> ``SDPBackend`` members."""
+    if names is None:
+        return None
+    from torch.nn.attention import SDPBackend  # a pybind enum: getattr, not subscript
+
+    backends = []
+    for name in names:
+        member = SDPA_BACKENDS.get(name.lower())
+        if member is None:
+            raise ValueError(
+                f"llm_conf.sdpa_backends: unknown backend {name!r}, "
+                f"choose from {sorted(SDPA_BACKENDS)}"
+            )
+        backends.append(getattr(SDPBackend, member))
+    return backends
+
+
+def configure_llm_forward(llm, llm_conf):
+    """Install the switches ``llm_conf`` asks for on ``llm.model.forward`` (nothing by default)."""
+    backends = resolve_sdpa_backends(llm_conf.get("sdpa_backends", None))
+    compile_requested = bool(llm_conf.get("torch_compile", False))
+    if backends is None and not compile_requested:
+        return
+
+    decoder = llm.model
+    forward = eager = decoder.forward
+
+    if compile_requested:
+        compiled = torch.compile(eager, dynamic=True)
+
+        def forward(*args, **kw):
+            # Per call, not at construction: the trainer builds the model on the CPU and moves it
+            # to the GPU afterwards. A 1-sequence batch would get its own specialised graph.
+            x = kw.get("inputs_embeds", kw.get("input_ids"))
+            if x is None and args:
+                x = args[0]
+            if x is None or x.device.type != "cuda" or x.shape[0] == 1:
+                return eager(*args, **kw)
+            return compiled(*args, **kw)
+
+    if backends is not None:
+        from torch.nn.attention import sdpa_kernel
+
+        inner = forward
+
+        def forward(*args, **kw):
+            with sdpa_kernel(backends):  # wraps the compiled call, so Dynamo traces under it
+                return inner(*args, **kw)
+
+    decoder.forward = forward

funasr/models/fun_asr_nano/model.py

@@ -24,6 +24,7 @@
 from .ctc import CTC
 from .checkpoint_utils import disable_incomplete_ctc, normalize_checkpoint_state
 from .device_utils import resolve_autocast_device_type
+from .llm_forward_opts import configure_llm_forward
 from .tools.utils import forced_align
 
 dtype_map = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}
@@ -127,6 +128,11 @@ def __init__(
 
         self.llm_dtype = llm_conf.get("llm_dtype", "fp32")
         self.llm = model.to(dtype_map[self.llm_dtype])
+        # Opt-in training-speed switches, both off by default (finetune.sh turns them on):
+        # llm_conf.sdpa_backends sets the process-wide SDPA backend flags while this forward
+        # runs and restores them on return; llm_conf.torch_compile compiles the decoder stack
+        # for inputs on a CUDA device (a CPU decoder stays eager). See llm_forward_opts.py.
+        configure_llm_forward(self.llm, llm_conf)
         llm_dim = model.get_input_embeddings().weight.shape[-1]
 
         # lora: inject LoRA adapters into the LLM target Linear layers

tests/test_fun_asr_nano_train_opts.py

@@ -0,0 +1,154 @@
+"""llm_conf.sdpa_backends / llm_conf.torch_compile on both copies of FunASRNano (the package
+class and the recipe's model.py, imported by path). CPU-only, no weights: tiny stand-ins."""
+
+import importlib.util
+import os
+import re
+import sys
+import types
+
+import pytest
+import torch
+import torch.nn as nn
+
+from funasr.models.fun_asr_nano import model as funasr_model
+from funasr.register import tables
+
+REPO = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+RECIPE_DIR = os.path.join(REPO, "examples", "industrial_data_pretraining", "fun_asr_nano")
+
+
+class _TinyDecoder(nn.Module):
+    """Stands in for the HF decoder stack (``llm.model``); records every eager call."""
+
+    def __init__(self, dim):
+        super().__init__()
+        self.proj = nn.Linear(dim, dim)
+        self.eager_calls = []
+
+    def forward(self, input_ids=None, inputs_embeds=None, **kwargs):
+        if not torch.compiler.is_compiling():
+            cudnn = torch.backends.cuda.cudnn_sdp_enabled()
+            self.eager_calls.append({"batch": int(inputs_embeds.shape[0]), "cudnn": cudnn})
+        return self.proj(inputs_embeds)
+
+
+class _TinyLLM(nn.Module):
+    def __init__(self, dim=8, vocab=16):
+        super().__init__()
+        self.model, self.embed = _TinyDecoder(dim), nn.Embedding(vocab, dim)
+
+    def get_input_embeddings(self):
+        return self.embed
+
+
+class _TinyEncoder(nn.Module):
+    def __init__(self, input_size=80, **kwargs):
+        super().__init__()
+        self.lin = nn.Linear(input_size, 4)
+
+    def output_size(self):
+        return 4
+
+
+class _TinyAdaptor(nn.Module):
+    def __init__(self, encoder_dim=4, llm_dim=8, **kwargs):
+        super().__init__()
+        self.lin = nn.Linear(encoder_dim, llm_dim)
+
+
+def _load_recipe_model():
+    previous = tables.model_classes.get("FunASRNano")
+    if RECIPE_DIR not in sys.path:
+        sys.path.insert(0, RECIPE_DIR)  # the recipe imports ``ctc`` and ``tools.utils`` bare
+    path = os.path.join(RECIPE_DIR, "model.py")
+    spec = importlib.util.spec_from_file_location("fun_asr_nano_recipe_model", path)
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module  # tables.register() calls inspect.getfile on the class
+    try:
+        spec.loader.exec_module(module)
+    except ImportError as exc:
+        sys.modules.pop(spec.name, None)
+        pytest.skip(f"recipe model.py not importable here: {exc}")
+    finally:
+        tables.model_classes["FunASRNano"] = previous  # the package class, registered at import
+    return module.FunASRNano
+
+
+@pytest.fixture(scope="module", params=["funasr", "recipe"])
+def model_class(request):
+    return funasr_model.FunASRNano if request.param == "funasr" else _load_recipe_model()
+
+
+@pytest.fixture
+def build(monkeypatch, model_class):
+    """``build(llm_conf) -> FunASRNano`` on the CPU, with the transformers loaders stubbed."""
+    fake = types.ModuleType("transformers")
+    fake.AutoConfig = types.SimpleNamespace(from_pretrained=lambda path, **kw: {})
+    fake.AutoModelForCausalLM = types.SimpleNamespace(from_config=lambda config, **kw: _TinyLLM())
+    monkeypatch.setattr(funasr_model, "AutoConfig", fake.AutoConfig)  # bound at import time
+    monkeypatch.setattr(funasr_model, "AutoModelForCausalLM", fake.AutoModelForCausalLM)
+    monkeypatch.setitem(sys.modules, "transformers", fake)  # the recipe imports in __init__
+    monkeypatch.setitem(tables.encoder_classes, "TinyEnc", _TinyEncoder)
+    monkeypatch.setitem(tables.adaptor_classes, "TinyAdp", _TinyAdaptor)
+    kw = dict(audio_encoder="TinyEnc", audio_adaptor="TinyAdp", llm="tiny")
+    return lambda c: model_class(audio_encoder_conf={}, audio_adaptor_conf={}, llm_conf=c, **kw)
+
+
+def _run(model, batch):
+    device = next(model.llm.parameters()).device
+    return model.llm.model.forward(inputs_embeds=torch.zeros(batch, 3, 8, device=device))
+
+
+def test_default_leaves_sdpa_flag_and_forward_alone(build):
+    flag = torch.backends.cuda.cudnn_sdp_enabled()
+    model = build({})
+    assert torch.backends.cuda.cudnn_sdp_enabled() == flag
+    assert "forward" not in model.llm.model.__dict__  # still the class method
+    _run(model, 2)
+    assert model.llm.model.eager_calls[-1]["cudnn"] == flag
+
+
+def test_sdpa_backends_disable_cudnn_only_inside_forward(build):
+    flag = torch.backends.cuda.cudnn_sdp_enabled()
+    model = build({"sdpa_backends": ["flash", "efficient", "math"]})
+    assert torch.backends.cuda.cudnn_sdp_enabled() == flag  # construction changed nothing
+    _run(model, 2)
+    assert model.llm.model.eager_calls[-1]["cudnn"] is False
+    assert torch.backends.cuda.cudnn_sdp_enabled() == flag  # restored on return
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a CUDA device")
+def test_torch_compile_wraps_cuda_llm_and_keeps_batch_one_eager(build):
+    model = build({"torch_compile": True}).cuda()  # built on the CPU and moved, as the trainer does
+    decoder = model.llm.model
+    assert "forward" in decoder.__dict__
+    _run(model, 1)
+    assert [c["batch"] for c in decoder.eager_calls] == [1]
+    out = _run(model, 2)  # compiled: the stand-in does not record under Dynamo
+    assert [c["batch"] for c in decoder.eager_calls] == [1]
+    torch.testing.assert_close(out, decoder.proj(torch.zeros(2, 3, 8, device=out.device)))
+
+
+def test_torch_compile_keeps_cpu_decoder_eager(build):
+    model = build({"torch_compile": True})
+    _run(model, 1)
+    _run(model, 4)
+    assert [c["batch"] for c in model.llm.model.eager_calls] == [1, 4]
+
+
+def test_recipe_declares_the_torch_floor_its_options_need():
+    """sdpa_backends goes through torch.nn.attention, absent before torch 2.3; if a recipe
+    turns it on, the example's requirements.txt has to say so."""
+    for recipe in ("finetune.sh", "lora_finetune.sh"):
+        path = os.path.join(RECIPE_DIR, recipe)
+        if not os.path.exists(path):
+            continue
+        with open(path) as f:
+            if "llm_conf.sdpa_backends" not in f.read():
+                continue
+        with open(os.path.join(RECIPE_DIR, "requirements.txt")) as f:
+            declared = re.search(r"^torch>=(\d+)\.(\d+)", f.read(), re.M)
+        assert declared, f"{recipe} sets sdpa_backends but requirements.txt pins no torch floor"
+        floor = (int(declared.group(1)), int(declared.group(2)))
+        assert floor >= (2, 3), f"{recipe} sets sdpa_backends; torch floor is {floor}, needs >= 2.3"