[docs] Loading behavior

docs/source/en/installation.md

@@ -168,7 +168,9 @@ After installation, you can configure the Transformers cache location or set up
 
 When you load a pretrained model with [`~PreTrainedModel.from_pretrained`], the model is downloaded from the Hub and locally cached.
 
-Every time you load a model, it checks whether the cached model is up-to-date. If it's the same, then the local model is loaded. If it's not the same, the newer model is downloaded and cached.
+If you pass a commit hash, Transformers uses the local cache for that commit's files and does not re-check the Hub for each one (including files that are known to be missing). This is the default behavior when no revision is specified.
+
+If you pass a branch or tag, Transformers contacts the Hub once at the start of the load to pick the commit, then uses the cache the same way for that commit's files.
 
 The default directory given by the shell environment variable `HF_HUB_CACHE` is `~/.cache/huggingface/hub`. On Windows, the default directory is `C:\Users\username\.cache\huggingface\hub`.
 
@@ -205,3 +207,5 @@ from transformers import LlamaForCausalLM
 
 model = LlamaForCausalLM.from_pretrained("./path/to/local/directory", local_files_only=True)
 ```
+
+Offline mode (or `local_files_only=True`) can still turn a branch or tag into a commit if an earlier online load saved that mapping in the cache. If the mapping was never saved, Transformers keeps the branch or tag you asked for and continues with the regular offline load. You get the same cache hits or missing-file errors as a normal offline load.

docs/source/en/model_sharing.md

@@ -63,6 +63,10 @@ model = AutoModel.from_pretrained(
 )
 ```
 
+Loading from a branch or tag pins every file in that call to one commit, then fetches every file from that commit.
+
+That pin lasts for the call only. For reproducibility across runs, pass an explicit commit hash or a tag you treat as fixed. Do not rely on `main` or another moving branch staying fixed between runs.
+
 Model repositories also support [gating](https://hf.co/docs/hub/models-gated) to control who can access a model. Gating is common for allowing a select group of users to preview a research model before it's made public.
 
 <div class="flex justify-center">