项目文件夹

文件
Alexandre Défossez (autodoc) aeffa9f270 api_docs
2025-03-13 16:07:03 +00:00

253 行
15 KiB
HTML

此文件含有不可见的 Unicode 字符
此文件含有人类无法区分的不可见的 Unicode 字符,但可以由计算机进行不同的处理。 如果您是想特意这样的,可以安全地忽略该警告。 使用 Escape 按钮显示他们。
此文件含有可能会与其他字符混淆的 Unicode 字符。 如果您是想特意这样的,可以安全地忽略该警告。 使用 Escape 按钮显示他们。
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, minimum-scale=1">
<meta name="generator" content="pdoc3 0.11.5">
<title>audiocraft.optim.fsdp API documentation</title>
<meta name="description" content="Wrapper around FSDP for more convenient use in the training loops.">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/13.0.0/sanitize.min.css" integrity="sha512-y1dtMcuvtTMJc1yPgEqF0ZjQbhnc/bFhyvIyVNb9Zk5mIGtqVaAB1Ttl28su8AvFMOY0EwRbAe+HCLqj6W7/KA==" crossorigin>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/13.0.0/typography.min.css" integrity="sha512-Y1DYSb995BAfxobCkKepB1BqJJTPrOp3zPL74AWFugHHmmdcvO+C48WLrUOlhGMc0QG7AE3f7gmvvcrmX2fDoA==" crossorigin>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/styles/default.min.css" crossorigin>
<style>:root{--highlight-color:#fe9}.flex{display:flex !important}body{line-height:1.5em}#content{padding:20px}#sidebar{padding:1.5em;overflow:hidden}#sidebar > *:last-child{margin-bottom:2cm}.http-server-breadcrumbs{font-size:130%;margin:0 0 15px 0}#footer{font-size:.75em;padding:5px 30px;border-top:1px solid #ddd;text-align:right}#footer p{margin:0 0 0 1em;display:inline-block}#footer p:last-child{margin-right:30px}h1,h2,h3,h4,h5{font-weight:300}h1{font-size:2.5em;line-height:1.1em}h2{font-size:1.75em;margin:2em 0 .50em 0}h3{font-size:1.4em;margin:1.6em 0 .7em 0}h4{margin:0;font-size:105%}h1:target,h2:target,h3:target,h4:target,h5:target,h6:target{background:var(--highlight-color);padding:.2em 0}a{color:#058;text-decoration:none;transition:color .2s ease-in-out}a:visited{color:#503}a:hover{color:#b62}.title code{font-weight:bold}h2[id^="header-"]{margin-top:2em}.ident{color:#900;font-weight:bold}pre code{font-size:.8em;line-height:1.4em;padding:1em;display:block}code{background:#f3f3f3;font-family:"DejaVu Sans Mono",monospace;padding:1px 4px;overflow-wrap:break-word}h1 code{background:transparent}pre{border-top:1px solid #ccc;border-bottom:1px solid #ccc;margin:1em 0}#http-server-module-list{display:flex;flex-flow:column}#http-server-module-list div{display:flex}#http-server-module-list dt{min-width:10%}#http-server-module-list p{margin-top:0}.toc ul,#index{list-style-type:none;margin:0;padding:0}#index code{background:transparent}#index h3{border-bottom:1px solid #ddd}#index ul{padding:0}#index h4{margin-top:.6em;font-weight:bold}@media (min-width:200ex){#index .two-column{column-count:2}}@media (min-width:300ex){#index .two-column{column-count:3}}dl{margin-bottom:2em}dl dl:last-child{margin-bottom:4em}dd{margin:0 0 1em 3em}#header-classes + dl > dd{margin-bottom:3em}dd dd{margin-left:2em}dd p{margin:10px 0}.name{background:#eee;font-size:.85em;padding:5px 10px;display:inline-block;min-width:40%}.name:hover{background:#e0e0e0}dt:target .name{background:var(--highlight-color)}.name > span:first-child{white-space:nowrap}.name.class > span:nth-child(2){margin-left:.4em}.inherited{color:#999;border-left:5px solid #eee;padding-left:1em}.inheritance em{font-style:normal;font-weight:bold}.desc h2{font-weight:400;font-size:1.25em}.desc h3{font-size:1em}.desc dt code{background:inherit}.source > summary,.git-link-div{color:#666;text-align:right;font-weight:400;font-size:.8em;text-transform:uppercase}.source summary > *{white-space:nowrap;cursor:pointer}.git-link{color:inherit;margin-left:1em}.source pre{max-height:500px;overflow:auto;margin:0}.source pre code{font-size:12px;overflow:visible;min-width:max-content}.hlist{list-style:none}.hlist li{display:inline}.hlist li:after{content:',\2002'}.hlist li:last-child:after{content:none}.hlist .hlist{display:inline;padding-left:1em}img{max-width:100%}td{padding:0 .5em}.admonition{padding:.1em 1em;margin:1em 0}.admonition-title{font-weight:bold}.admonition.note,.admonition.info,.admonition.important{background:#aef}.admonition.todo,.admonition.versionadded,.admonition.tip,.admonition.hint{background:#dfd}.admonition.warning,.admonition.versionchanged,.admonition.deprecated{background:#fd4}.admonition.error,.admonition.danger,.admonition.caution{background:lightpink}</style>
<style media="screen and (min-width: 700px)">@media screen and (min-width:700px){#sidebar{width:30%;height:100vh;overflow:auto;position:sticky;top:0}#content{width:70%;max-width:100ch;padding:3em 4em;border-left:1px solid #ddd}pre code{font-size:1em}.name{font-size:1em}main{display:flex;flex-direction:row-reverse;justify-content:flex-end}.toc ul ul,#index ul ul{padding-left:1em}.toc > ul > li{margin-top:.5em}}</style>
<style media="print">@media print{#sidebar h1{page-break-before:always}.source{display:none}}@media print{*{background:transparent !important;color:#000 !important;box-shadow:none !important;text-shadow:none !important}a[href]:after{content:" (" attr(href) ")";font-size:90%}a[href][title]:after{content:none}abbr[title]:after{content:" (" attr(title) ")"}.ir a:after,a[href^="javascript:"]:after,a[href^="#"]:after{content:""}pre,blockquote{border:1px solid #999;page-break-inside:avoid}thead{display:table-header-group}tr,img{page-break-inside:avoid}img{max-width:100% !important}@page{margin:0.5cm}p,h2,h3{orphans:3;widows:3}h1,h2,h3,h4,h5,h6{page-break-after:avoid}}</style>
<script defer src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/highlight.min.js" integrity="sha512-D9gUyxqja7hBtkWpPWGt9wfbfaMGVt9gnyCvYa+jojwwPHLCzUm5i8rpk7vD7wNee9bA35eYIjobYPaQuKS1MQ==" crossorigin></script>
<script>window.addEventListener('DOMContentLoaded', () => {
hljs.configure({languages: ['bash', 'css', 'diff', 'graphql', 'ini', 'javascript', 'json', 'plaintext', 'python', 'python-repl', 'rust', 'shell', 'sql', 'typescript', 'xml', 'yaml']});
hljs.highlightAll();
/* Collapse source docstrings */
setTimeout(() => {
[...document.querySelectorAll('.hljs.language-python > .hljs-string')]
.filter(el => el.innerHTML.length > 200 && ['"""', "'''"].includes(el.innerHTML.substring(0, 3)))
.forEach(el => {
let d = document.createElement('details');
d.classList.add('hljs-string');
d.innerHTML = '<summary>"""</summary>' + el.innerHTML.substring(3);
el.replaceWith(d);
});
}, 100);
})</script>
</head>
<body>
<main>
<article id="content">
<header>
<h1 class="title">Module <code>audiocraft.optim.fsdp</code></h1>
</header>
<section id="section-intro">
<p>Wrapper around FSDP for more convenient use in the training loops.</p>
</section>
<section>
</section>
<section>
</section>
<section>
<h2 class="section-title" id="header-functions">Functions</h2>
<dl>
<dt id="audiocraft.optim.fsdp.is_fsdp_used"><code class="name flex">
<span>def <span class="ident">is_fsdp_used</span></span>(<span>) > bool</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def is_fsdp_used() -&gt; bool:
&#34;&#34;&#34;Return whether we are using FSDP.&#34;&#34;&#34;
# A bit of a hack but should work from anywhere.
if dora.is_xp():
cfg = dora.get_xp().cfg
if hasattr(cfg, &#39;fsdp&#39;):
return cfg.fsdp.use
return False</code></pre>
</details>
<div class="desc"><p>Return whether we are using FSDP.</p></div>
</dd>
<dt id="audiocraft.optim.fsdp.is_sharded_tensor"><code class="name flex">
<span>def <span class="ident">is_sharded_tensor</span></span>(<span>x: Any) > bool</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def is_sharded_tensor(x: tp.Any) -&gt; bool:
return isinstance(x, ShardedTensor)</code></pre>
</details>
<div class="desc"></div>
</dd>
<dt id="audiocraft.optim.fsdp.purge_fsdp"><code class="name flex">
<span>def <span class="ident">purge_fsdp</span></span>(<span>model: torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel)</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def purge_fsdp(model: FSDP):
&#34;&#34;&#34;Purge the FSDP cached shard inside the model. This should
allow setting the best state or switching to the EMA.
&#34;&#34;&#34;
from torch.distributed.fsdp._runtime_utils import _reshard # type: ignore
for module in FSDP.fsdp_modules(model):
if hasattr(module, &#34;_handles&#34;):
# support for FSDP with torch&lt;2.1.0
handles = module._handles
if not handles:
continue
handle = handles[0]
unsharded_flat_param = handle._get_padded_unsharded_flat_param()
storage_size: int = unsharded_flat_param._typed_storage()._size() # type: ignore
if storage_size == 0:
continue
true_list = [True for h in handles]
_reshard(module, handles, true_list)
else:
handle = module._handle
if not handle:
continue
unsharded_flat_param = handle._get_padded_unsharded_flat_param()
storage_size: int = unsharded_flat_param._typed_storage()._size() # type: ignore
if storage_size == 0:
continue
_reshard(module, handle, True)</code></pre>
</details>
<div class="desc"><p>Purge the FSDP cached shard inside the model. This should
allow setting the best state or switching to the EMA.</p></div>
</dd>
<dt id="audiocraft.optim.fsdp.switch_to_full_state_dict"><code class="name flex">
<span>def <span class="ident">switch_to_full_state_dict</span></span>(<span>models: List[torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel])</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@contextmanager
def switch_to_full_state_dict(models: tp.List[FSDP]):
# Another bug in FSDP makes it that we cannot use the `state_dict_type` API,
# so let&#39;s do thing manually.
for model in models:
FSDP.set_state_dict_type( # type: ignore
model, StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True))
try:
yield
finally:
for model in models:
FSDP.set_state_dict_type(model, StateDictType.LOCAL_STATE_DICT) # type: ignore</code></pre>
</details>
<div class="desc"></div>
</dd>
<dt id="audiocraft.optim.fsdp.wrap_with_fsdp"><code class="name flex">
<span>def <span class="ident">wrap_with_fsdp</span></span>(<span>cfg,<br>model: torch.nn.modules.module.Module,<br>block_classes: Set[Type] | None = None) > torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def wrap_with_fsdp(cfg, model: torch.nn.Module,
block_classes: tp.Optional[tp.Set[tp.Type]] = None) -&gt; FSDP:
&#34;&#34;&#34;Wraps a model with FSDP.&#34;&#34;&#34;
# Some of the typing is disabled until this gets integrated
# into the stable version of PyTorch.
from torch.distributed.fsdp.wrap import ModuleWrapPolicy # type: ignore
# we import this here to prevent circular import.
from ..modules.transformer import StreamingTransformerLayer
from ..modules.conditioners import ConditioningProvider
_fix_post_backward_hook()
assert cfg.use
sharding_strategy_dict = {
&#34;no_shard&#34;: ShardingStrategy.NO_SHARD,
&#34;shard_grad_op&#34;: ShardingStrategy.SHARD_GRAD_OP,
&#34;full_shard&#34;: ShardingStrategy.FULL_SHARD,
}
dtype_dict = {
&#34;float32&#34;: torch.float32,
&#34;float16&#34;: torch.float16,
&#34;bfloat16&#34;: torch.bfloat16,
}
mixed_precision_config = MixedPrecision(
param_dtype=dtype_dict[cfg.param_dtype],
reduce_dtype=dtype_dict[cfg.reduce_dtype],
buffer_dtype=dtype_dict[cfg.buffer_dtype],
)
sharding_strategy_config = sharding_strategy_dict[cfg.sharding_strategy]
# The following is going to require being a bit smart
# when doing LM, because this would flush the weights for every time step
# during generation. One possiblity is to use hybrid sharding:
# See: https://pytorch.org/docs/master/fsdp.html#torch.distributed.fsdp.ShardingStrategy
assert sharding_strategy_config != ShardingStrategy.FULL_SHARD, \
&#34;Not supported at the moment, requires a bit more work.&#34;
local_rank = dora.distrib.get_distrib_spec().local_rank
assert local_rank &lt; torch.cuda.device_count(), &#34;Please upgrade Dora!&#34;
auto_wrap_policy = None
if block_classes is None:
block_classes = {StreamingTransformerLayer, ConditioningProvider}
if cfg.per_block:
auto_wrap_policy = ModuleWrapPolicy(block_classes)
wrapped = _FSDPFixStateDict(
model,
sharding_strategy=sharding_strategy_config,
mixed_precision=mixed_precision_config,
device_id=local_rank,
sync_module_states=True,
use_orig_params=True,
auto_wrap_policy=auto_wrap_policy,
) # type: ignore
FSDP.set_state_dict_type(wrapped, StateDictType.LOCAL_STATE_DICT) # type: ignore
# Let the wrapped model know about the wrapping!
# We use __dict__ to avoid it going into the state dict.
# This is a bit dirty, but needed during generation, as otherwise
# the wrapped model would call itself and bypass FSDP.
for module in FSDP.fsdp_modules(wrapped):
original = module._fsdp_wrapped_module
original.__dict__[&#39;_fsdp&#39;] = module
return wrapped</code></pre>
</details>
<div class="desc"><p>Wraps a model with FSDP.</p></div>
</dd>
</dl>
</section>
<section>
</section>
</article>
<nav id="sidebar">
<div class="toc">
<ul></ul>
</div>
<ul id="index">
<li><h3>Super-module</h3>
<ul>
<li><code><a title="audiocraft.optim" href="index.html">audiocraft.optim</a></code></li>
</ul>
</li>
<li><h3><a href="#header-functions">Functions</a></h3>
<ul class="">
<li><code><a title="audiocraft.optim.fsdp.is_fsdp_used" href="#audiocraft.optim.fsdp.is_fsdp_used">is_fsdp_used</a></code></li>
<li><code><a title="audiocraft.optim.fsdp.is_sharded_tensor" href="#audiocraft.optim.fsdp.is_sharded_tensor">is_sharded_tensor</a></code></li>
<li><code><a title="audiocraft.optim.fsdp.purge_fsdp" href="#audiocraft.optim.fsdp.purge_fsdp">purge_fsdp</a></code></li>
<li><code><a title="audiocraft.optim.fsdp.switch_to_full_state_dict" href="#audiocraft.optim.fsdp.switch_to_full_state_dict">switch_to_full_state_dict</a></code></li>
<li><code><a title="audiocraft.optim.fsdp.wrap_with_fsdp" href="#audiocraft.optim.fsdp.wrap_with_fsdp">wrap_with_fsdp</a></code></li>
</ul>
</li>
</ul>
</nav>
</main>
<footer id="footer">
<p>Generated by <a href="https://pdoc3.github.io/pdoc" title="pdoc: Python API documentation generator"><cite>pdoc</cite> 0.11.5</a>.</p>
</footer>
</body>
</html>