accelerate/pr_4116/en/_app/immutable/nodes/26.CMdcbVCg.js

import"../chunks/DsnmJJEf.js";import{i as O,h as E,C as k,H as a,D as r,E as D,s as S}from"../chunks/ClLAY5C0.js";import{p as A,o as H,s as e,f as B,a as w,b as C,d as l,c as x,n as t,r as o}from"../chunks/BhevU81Y.js";const I='{"title":"FP8","local":"fp8","sections":[{"title":"FP8RecipeKwargs","local":"accelerate.utils.FP8RecipeKwargs","sections":[],"depth":2},{"title":"convert_model","local":"accelerate.utils.convert_model","sections":[],"depth":2},{"title":"has_transformer_engine_layers","local":"accelerate.utils.has_transformer_engine_layers","sections":[],"depth":2},{"title":"contextual_fp8_autocast","local":"accelerate.utils.contextual_fp8_autocast","sections":[],"depth":2},{"title":"apply_fp8_autowrap","local":"accelerate.utils.apply_fp8_autowrap","sections":[],"depth":2}],"depth":1}';var j=x('<meta name="hf:doc:metadata"/>'),z=x(`<p></p> <!> <!> <p>Below are functions and classes relative to the underlying FP8 implementation</p> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Deprecated. Please use one of the proper FP8 recipe kwargs classes such as <code>TERecipeKwargs</code> or <code>MSAMPRecipeKwargs</code> instead.</p></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Recursively converts the linear and layernorm layers of a model to their <code>transformers_engine</code> counterpart.</p></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Returns whether a given model has some <code>transformer_engine</code> layer or not.</p></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Wrapper for a model’s forward method to apply FP8 autocast. Is context aware, meaning that by default it will
disable FP8 autocast during eval mode, which is generally better for more accurate metrics.</p></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Applies FP8 context manager to the model’s forward method</p></div> <!> <p></p>`,1);function Y(P,F){A(F,!1),H(()=>{new URLSearchParams(window.location.search).get("fw")}),O();var d=z();E("ykje5o",b=>{var y=j();S(y,"content",I),w(b,y)});var m=e(B(d),2);k(m,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var _=e(m,2);a(_,{title:"FP8",local:"fp8",headingTag:"h1"});var u=e(_,4);a(u,{title:"FP8RecipeKwargs",local:"accelerate.utils.FP8RecipeKwargs",headingTag:"h2"});var n=e(u,2),L=l(n);r(L,{name:"class accelerate.utils.FP8RecipeKwargs",anchor:"accelerate.utils.FP8RecipeKwargs",source:"https://github.com/huggingface/accelerate/blob/vr_4116/src/accelerate/utils/dataclasses.py#L457",parameters:[{name:"opt_level",val:": typing.Literal['O1', 'O2'] = None"},{name:"use_autocast_during_eval",val:": typing.Optional[bool] = None"},{name:"margin",val:": typing.Optional[int] = None"},{name:"interval",val:": typing.Optional[int] = None"},{name:"fp8_format",val:": typing.Literal['HYBRID', 'E4M3', 'E5M2'] = None"},{name:"amax_history_len",val:": typing.Optional[int] = None"},{name:"amax_compute_algo",val:": typing.Literal['max', 'most_recent'] = None"},{name:"override_linear_precision",val:": tuple = None"},{name:"use_mxfp8_block_scaling",val:": typing.Optional[bool] = None"},{name:"backend",val:": typing.Literal['MSAMP', 'TE'] = None"}]}),t(2),o(n);var g=e(n,2);a(g,{title:"convert_model",local:"accelerate.utils.convert_model",headingTag:"h2"});var c=e(g,2),R=l(c);r(R,{name:"accelerate.utils.convert_model",anchor:"accelerate.utils.convert_model",source:"https://github.com/huggingface/accelerate/blob/vr_4116/src/accelerate/utils/transformer_engine.py#L26",parameters:[{name:"model",val:""},{name:"to_transformer_engine",val:" = True"},{name:"_convert_linear",val:" = True"},{name:"_convert_ln",val:" = True"}]}),t(2),o(c);var v=e(c,2);a(v,{title:"has_transformer_engine_layers",local:"accelerate.utils.has_transformer_engine_layers",headingTag:"h2"});var s=e(v,2),T=l(s);r(T,{name:"accelerate.utils.has_transformer_engine_layers",anchor:"accelerate.utils.has_transformer_engine_layers",source:"https://github.com/huggingface/accelerate/blob/vr_4116/src/accelerate/utils/transformer_engine.py#L95",parameters:[{name:"model",val:""}]}),t(2),o(s);var h=e(s,2);a(h,{title:"contextual_fp8_autocast",local:"accelerate.utils.contextual_fp8_autocast",headingTag:"h2"});var i=e(h,2),N=l(i);r(N,{name:"accelerate.utils.contextual_fp8_autocast",anchor:"accelerate.utils.contextual_fp8_autocast",source:"https://github.com/huggingface/accelerate/blob/vr_4116/src/accelerate/utils/transformer_engine.py#L118",parameters:[{name:"model_forward",val:""},{name:"fp8_recipe",val:""},{name:"use_during_eval",val:" = False"}]}),t(2),o(i);var f=e(i,2);a(f,{title:"apply_fp8_autowrap",local:"accelerate.utils.apply_fp8_autowrap",headingTag:"h2"});var p=e(f,2),M=l(p);r(M,{name:"accelerate.utils.apply_fp8_autowrap",anchor:"accelerate.utils.apply_fp8_autowrap",source:"https://github.com/huggingface/accelerate/blob/vr_4116/src/accelerate/utils/transformer_engine.py#L142",parameters:[{name:"model",val:""},{name:"fp8_recipe_handler",val:""}]}),t(2),o(p);var K=e(p,2);D(K,{source:"https://github.com/huggingface/accelerate/blob/main/docs/source/package_reference/fp8.md"}),t(2),w(P,d),C()}export{Y as component};