Sync upstream transformers main into OMY

.github/workflows/ssh-runner.yml

@@ -161,7 +161,7 @@ jobs:
           echo 'source /root/.env_setup' >> /root/.bashrc
 
       - name: Tailscale # In order to be able to SSH when a test fails
-        uses: huggingface/tailscale-action@7d53c9737e53934c30290b5524d1c9b4a7c98c8a  # main
+        uses: huggingface/tailscale-action@9d1a67c46875a956cf2d1975473da978a970a2a9  # main
         with:
           authkey: ${{ secrets.TAILSCALE_SSH_AUTHKEY }}
           slackChannel: ${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}

docs/source/en/_toctree.yml

@@ -667,10 +667,10 @@
         title: GLM-4.5, GLM-4.6, GLM-4.7
       - local: model_doc/glm4_moe_lite
         title: GLM-4.7-Flash
+      - local: model_doc/glm_moe_dsa
+        title: GLM-5, GLM-5.1, GLM-5.2, GLM-5.3
       - local: model_doc/glm_image
         title: GLM-Image
-      - local: model_doc/glm_moe_dsa
-        title: GlmMoeDsa
       - local: model_doc/openai-gpt
         title: GPT
       - local: model_doc/gpt_neo

docs/source/en/continuous_batching.md

@@ -195,7 +195,10 @@ for chunk in manager.request_id_iter(request_id="streamed"):
 
 ## ContinuousBatchingConfig
 
-[`ContinuousBatchingConfig`] controls the KV cache, scheduling, CUDA graphs, memory usage, and more. Pass it alongside [`GenerationConfig`] to customize continuous batching.
+[`ContinuousBatchingConfig`] controls the KV cache, scheduling, CUDA graphs, memory usage, and more. Pass it to `generate_batch` or [`~ContinuousMixin.init_continuous_batching`] with the `continuous_batching_config` argument, separately from [`GenerationConfig`]. The two configs describe different things. [`GenerationConfig`] is model-centered and holds sampling and stopping parameters, while [`ContinuousBatchingConfig`] is hardware-centered and holds memory and scheduling parameters.
+
+> [!WARNING]
+> Setting `continuous_batching_config` on a [`GenerationConfig`] is deprecated, emits a `FutureWarning`, and will be removed in v5.19.
 
 By default, `max_batch_tokens` is `8192`, bounded by available GPU memory and never below `256`, while `num_blocks` fills the remaining memory. Use the table below to help you pick the appropriate features.
 
@@ -358,6 +361,8 @@ The fast path relies on the `flash_attn_with_kvcache` kernel, which is available
 
 For any other combination, or when the kernel can't be imported, the manager falls back to the varlen path. It logs a warning only when you set `max_blocks_per_request` explicitly.
 
+Sliding window attention doesn't support block tables, so the cache forces `max_blocks_per_request` to `0` for any model with sliding window layers, regardless of the attention implementation. If you set a nonzero value, it's overridden and the cache logs `Sliding window attention groups detected: disabling block table support.`
+
 ### CPU offloading
 
 CPU offloading copies evicted KV cache blocks to a pre-allocated pinned CPU buffer when the GPU KV cache is full. After cache space becomes available, the manager copies the blocks back to the GPU and resumes the request without recomputing its prompt and generated tokens.
@@ -470,7 +475,7 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 ```
 
-Prefix caching is disabled automatically when sliding window attention is active.
+Prefix caching and the [decode fast path](#decode-fast-path) are disabled automatically when sliding window attention is active.
 
 ## Next steps
 

docs/source/en/model_doc/glm_moe_dsa.md

@@ -25,11 +25,21 @@ limitations under the License.
     </div>
 </div>
 
-# GLM-5, GLM-5.1, GLM-5.2
+# GLM-5, GLM-5.1, GLM-5.2, GLM-5.3
 
 ## Overview
 
-**GLM-5**, **GLM-5.1** and **GLM-5.2** language model use this class. The implementation in transformers does not include an MTP layer.
+**GLM-5**, **GLM-5.1**, **GLM-5.2** and **GLM-5.3** language model use this class. The implementation in transformers does not include an MTP layer.
+
+### GLM-5.3
+
+GLM-5.3 uses the same base model as GLM-5.2 — every gain comes from post-training. Compared with GLM-5.2, it is much better at complex coding and long-horizon tasks:
+
++ Stronger Coding: GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench. It also achieve open-source SOTA on public benchmarks including Terminal Bench 3.0 and
+  Agents' Last Exam.
++ Emergent Cyber Capability: As we scaled post-training, cyber capability developed faster than we expected. GLM-5.3 is state of the art on CyberGym for vulnerability discovery, and its gains are largest further up the exploitation chain, where it more than doubles GLM-5.2 on exploitation benchmarks.
+
+![bench_53](https://raw.githubusercontent.com/zai-org/GLM-5/refs/heads/main/resources/bench_53_2.png)
 
 ### GLM-5.2