Llama-3.1-405B-Instruct-autofp8

Llama-3.1-405B-Instruct quantized to fp8 using AutoFP8.

This model has tensorwise fp8 scale factors.

AutoFP8 is old, I had to apply the following patch to run:

diff --git a/auto_fp8/modeling.py b/auto_fp8/modeling.py
index 04a9e71..9fdacda 100644
--- a/auto_fp8/modeling.py
+++ b/auto_fp8/modeling.py
@@ -58,10 +58,10 @@ class AutoFP8ForCausalLM:
         # Parameters related to loading from Hugging Face Hub
         cache_dir = model_init_kwargs.pop("cache_dir", None)
         force_download = model_init_kwargs.pop("force_download", False)
-        resume_download = model_init_kwargs.pop("resume_download", False)
+        #resume_download = model_init_kwargs.pop("resume_download", False)
         proxies = model_init_kwargs.pop("proxies", None)
         local_files_only = model_init_kwargs.pop("local_files_only", False)
-        use_auth_token = model_init_kwargs.pop("use_auth_token", None)
+        #use_auth_token = model_init_kwargs.pop("use_auth_token", None)
         revision = model_init_kwargs.pop("revision", None)
         subfolder = model_init_kwargs.pop("subfolder", "")
         commit_hash = model_init_kwargs.pop("_commit_hash", None)
@@ -70,9 +70,9 @@ class AutoFP8ForCausalLM:
             "cache_dir": cache_dir,
             "force_download": force_download,
             "proxies": proxies,
-            "resume_download": resume_download,
+            #"resume_download": resume_download,
             "local_files_only": local_files_only,
-            "use_auth_token": use_auth_token,
+            #"use_auth_token": use_auth_token,
             "revision": revision,
             "subfolder": subfolder,
             "_commit_hash": commit_hash,

The tokenizer_class has also been manually updated to use PreTrainedTokenizerFast