add glm support + patch (#3329) [skip ci]

* add glm support + patch

* lint

* lint

* Update examples/glm4/glm-4-6v-flash-qlora.yaml

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* Update examples/glm4/glm-4-6v-flash-qlora.yaml

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* Update src/axolotl/processing_strategies.py

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* patch removed

* lint

* lint2

* docs + rename

* rmv moe

* docs

* removed processor

* sdpa T_T"

* ddp_find_unused_parameters: true

* muti gpu yaml tested both

* muti gpu yaml tested both

* Update examples/glm46v/README.md

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* Update examples/glm46v/README.md

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* Update examples/glm46v/README.md

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

* rmv text only section + v5 comments

* rename

---------

Co-authored-by: Ved <ved.work2024@gmail.com>
Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

docs/multimodal.qmd

@@ -19,6 +19,7 @@ format:
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [GLM-4.6V](#sec-glm-4-6v)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
 - [Intern-VL](#sec-intern-vl)
@@ -183,6 +184,18 @@ base_model: Qwen/Qwen3-VL-4B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### GLM-4.6V {#sec-glm-4-6v}
+
+Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.
+
+```yaml
+# GLM-4.6V (106B MoE version)
+base_model: zai-org/GLM-4.6V
+
+# OR GLM-4.6V-Flash (9B version)
+base_model: zai-org/GLM-4.6V-Flash
+```
+
 ### SmolVLM2 {#sec-smolvlm2}
 
 ::: {.callout-tip}

examples/glm46v/README.md

@@ -0,0 +1,44 @@
+# Finetune GLM-4.6V with Axolotl
+
+GLM-4.6V is a family of vision-language models from ZhipuAI found on [HuggingFace](https://huggingface.co/zai-org/GLM-4.6V). This guide shows how to fine-tune it with Axolotl for vision-language tasks.
+
+
+
+## Getting started
+
+1. Install Axolotl from source following the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+
+3. Run the fine-tuning:
+
+    glm-4-6v-flash(9B)
+    ```bash
+    axolotl train examples/glm46v/glm-4-6v-flash-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+## Tips
+
+- Vision datasets should follow the format described in the [multimodal docs](https://docs.axolotl.ai/docs/multimodal.html#dataset-format)
+- You can run a **full finetuning** by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset in the [dataset loading docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+
+## Supported Models
+
+- **GLM-4.6V**: Full vision-language model (`zai-org/GLM-4.6V`)
+- **GLM-4.6V-Flash**: Faster variant (`zai-org/GLM-4.6V-Flash`)
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [ZhipuAI GLM-4.6V](https://huggingface.co/zai-org/GLM-4.6V)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/glm46v/glm-4-6v-flash-ddp.yaml

@@ -0,0 +1,53 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+ddp_find_unused_parameters: true
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+sequence_len: 2048
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

examples/glm46v/glm-4-6v-flash-qlora.yaml

@@ -0,0 +1,50 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+sequence_len: 2048
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

src/axolotl/processing_strategies.py

@@ -485,6 +485,58 @@ class InternVLProcessingStrategy(ProcessingStrategy):
         return labels
 
 
+class Glm4vProcessingStrategy(ProcessingStrategy):
+    """Processing Strategy class for GLM4V and GLM4V-MoE vision models."""
+
+    def __init__(
+        self,
+        processor: ProcessorMixin,
+        chat_template: Optional[str] = None,
+        image_size: int | tuple[int, int] | None = None,
+        image_resize_algorithm: Resampling | None = None,
+    ):
+        super().__init__(processor, chat_template, image_size, image_resize_algorithm)
+
+        self.tokenizer = getattr(processor, "tokenizer", processor)
+
+        self.image_token = "<|image|>"  # nosec
+        self.begin_image_token = "<|begin_of_image|>"  # nosec
+        self.end_image_token = "<|end_of_image|>"  # nosec
+        self.video_token = "<|video|>"  # nosec
+        self.begin_video_token = "<|begin_of_video|>"  # nosec
+        self.end_video_token = "<|end_of_video|>"  # nosec
+
+        self.image_token_id = self.tokenizer.convert_tokens_to_ids(self.image_token)
+        self.begin_image_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.begin_image_token
+        )
+        self.end_image_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.end_image_token
+        )
+        self.video_token_id = self.tokenizer.convert_tokens_to_ids(self.video_token)
+        self.begin_video_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.begin_video_token
+        )
+        self.end_video_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.end_video_token
+        )
+
+    def process_labels(self, input_ids):
+        labels = input_ids.clone()
+
+        labels[labels == self.tokenizer.pad_token_id] = -100
+
+        labels[labels == self.image_token_id] = -100
+        labels[labels == self.begin_image_token_id] = -100
+        labels[labels == self.end_image_token_id] = -100
+
+        labels[labels == self.video_token_id] = -100
+        labels[labels == self.begin_video_token_id] = -100
+        labels[labels == self.end_video_token_id] = -100
+
+        return labels
+
+
 def get_processing_strategy(
     processor: ProcessorMixin,
     chat_template,
@@ -501,10 +553,10 @@ def get_processing_strategy(
         "image_resize_algorithm": image_resize_algorithm,
     }
 
-    if chat_template_type in [None, "tokenizer_default"] and hasattr(
+    if chat_template_type in [None, "tokenizer_default"]:
-        processor.tokenizer, "chat_template"
+        tokenizer = getattr(processor, "tokenizer", processor)
-    ):
+        if hasattr(tokenizer, "chat_template"):
-        processing_kwargs["chat_template"] = processor.tokenizer.chat_template
+            processing_kwargs["chat_template"] = tokenizer.chat_template
 
     if chat_template_type == "qwen2_vl":
         return Qwen2VLProcessingStrategy(
@@ -533,6 +585,15 @@ def get_processing_strategy(
         return Mistral3ProcessingStrategy(
             **processing_kwargs,
         )
+    try:
+        from transformers.models.glm46v.processing_glm46v import Glm46VProcessor
+
+        if isinstance(processor, Glm46VProcessor):
+            return Glm4vProcessingStrategy(
+                **processing_kwargs,
+            )
+    except ImportError:
+        pass
 
     if isinstance(processor, InternVLProcessor):
         return InternVLProcessingStrategy(