feat(glm5-next): add GLM-5.3-Flash training support

docs/fern/versions/nightly.yml

@@ -268,6 +268,9 @@ navigation:
           - page: "MiniMax-M3"
             path: ../../model-coverage/vlm/minimax/minimax-m3.mdx
             slug: minimax-m3
+          - page: "GLM-5.3-Flash"
+            path: ../../model-coverage/vlm/thudm/glm-5-3-flash.mdx
+            slug: glm-5-3-flash
       - section: "Multimodal"
         slug: multimodal
         contents:

docs/model-coverage/vlm/thudm/glm-5-3-flash.mdx

@@ -0,0 +1,41 @@
+---
+title: "GLM-5.3-Flash"
+description: "Fine-tune the GLM-5.3-Flash mixture-of-experts vision-language model with packed context and expert parallelism."
+slug: model-coverage/vision-language-models/thudm/glm-5-3-flash
+---
+
+[GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) is a mixture-of-experts vision-language model with a hybrid Kimi Delta Attention and Dynamic Sparse Attention language backbone.
+
+<Info>
+
+| | |
+|---|---|
+| **Task** | Image-Text-to-Text |
+| **Architecture** | `Glm5NextForConditionalGeneration` |
+| **Language Module** | Hybrid KDA / KPool-DSA MoE decoder |
+| **Training Precision** | BF16 after FP8 checkpoint dequantization |
+| **HF Org** | [zai-org](https://huggingface.co/zai-org) |
+
+</Info>
+
+## Supported Training Path
+
+NeMo AutoModel provides a native configuration, image processor, vision tower, language model, and Hugging Face state-dict adapter for GLM-5.3-Flash. The supported full-model initialization path loads the base checkpoint through distributed checkpointing; single-GPU full-checkpoint loading is not supported.
+
+The current onboarding supports image training. Video inputs, tensor parallelism, and pipeline parallelism are not enabled for this model. The validated recipe uses FSDP2 with expert parallelism, contiguous packed context parallelism, and HybridEP dispatch.
+
+## Attention Backends
+
+- KDA layers use Flash Linear Attention kernels.
+- Sparse DSA layers support the SDPA reference path.
+- On SM90 or later, `backend.attn: cudnn` uses FlashMLA forward with cuDNN sparse-attention backward. This optional path requires compatible FlashMLA and cuDNN Frontend installations.
+
+## Example Recipe
+
+- [Full SFT — MedPix, packed 2K, EP72 + CP2](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/vlm_finetune/glm5_next/glm5_3_flash_medpix_packed2k_ep72_cp2_100steps.yaml)
+
+The recipe is sized for 9 nodes with 8 GPUs per node. It uses local batch size 1 and four gradient-accumulation microsteps to form a global batch of 144 packed samples.
+
+## Hugging Face Model Card
+
+- [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash)

examples/vlm_finetune/glm5_next/glm5_3_flash_medpix_packed2k_ep72_cp2_100steps.yaml

@@ -0,0 +1,155 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Validated 9-node/72-GPU packed MedPix run for GLM-5.3-Flash. CP2 gives DP36;
+# EP72 spans the dp(36) x cp(2) mesh and stores four of 288 routed experts per
+# GPU. Four accumulation microsteps form each 144-pack optimizer step. TP and PP
+# remain disabled.
+
+recipe: FinetuneRecipeForVLM
+seed: 1234
+
+step_scheduler:
+  global_batch_size: 144
+  local_batch_size: 1
+  ckpt_every_steps: 1000
+  val_every_steps: 1000
+  log_remote_every_steps: 1
+  gc_every_steps: 5
+  num_epochs: 4
+  max_steps: 100
+
+dist_env:
+  backend: nccl
+  timeout_minutes: 120
+
+rng:
+  _target_: nemo_automodel.components.training.rng.StatefulRNG
+  seed: 1234
+  ranked: true
+  deterministic: true
+
+model:
+  _target_: nemo_automodel.NeMoAutoModelForImageTextToText.from_pretrained
+  pretrained_model_name_or_path: zai-org/GLM-5.3-Flash
+  torch_dtype: bfloat16
+  attn_implementation: sdpa
+  use_liger_kernel: false
+  use_sdpa_patching: false
+  text_config:
+    output_hidden_states: true
+    num_nextn_predict_layers: 0
+  backend:
+    _target_: nemo_automodel.components.models.common.BackendConfig
+    # KDA layers remain on FLA; this selects cuDNN/FlashMLA for sparse MLA layers.
+    attn: cudnn
+    linear: torch
+    rms_norm: torch_fp32
+    experts: torch_mm
+    dispatcher: hybridep
+    rope_fusion: false
+    gate_precision: float32
+    fake_balanced_gate: false
+    enable_hf_state_dict_adapter: true
+    enable_fsdp_optimizations: true
+
+processor:
+  _target_: nemo_automodel.components.models.glm5_next.processing.build_glm5_next_processor
+  pretrained_model_name_or_path: zai-org/GLM-5.3-Flash
+  trust_remote_code: false
+
+checkpoint:
+  enabled: false
+  checkpoint_dir: checkpoints/glm5_3_flash_medpix_packed2k_ep72_cp2_100steps/
+  model_save_format: safetensors
+  save_consolidated: false
+  dequantize_base_checkpoint: true
+
+distributed:
+  strategy: fsdp2
+  tp_size: 1
+  cp_size: 2
+  pp_size: 1
+  ep_size: 72
+  sequence_parallel: false
+  activation_checkpointing: true
+  # Avoid retaining a full accumulated gradient set across four microsteps.
+  defer_fsdp_grad_sync: false
+  moe:
+    reshard_after_forward: false
+    wrap_outer_model: true
+    ignore_router_for_ac: true
+
+freeze_config:
+  freeze_embeddings: true
+  freeze_vision_tower: true
+  freeze_audio_tower: true
+  freeze_language_model: false
+
+loss_fn:
+  _target_: nemo_automodel.components.loss.linear_ce.FusedLinearCrossEntropy
+
+dataset:
+  _target_: nemo_automodel.components.datasets.vlm.datasets.make_medpix_dataset
+  path_or_dataset: mmoukouba/MedPix-VQA
+  split: train
+
+packed_sequence:
+  pretokenize: true
+  max_length: 2048
+  pack_size: 2048
+  collate_max_length: 2048
+  packing_ratio: 0.9
+  drop_long_samples: true
+  balance_media_tokens: true
+  packing_format: neat
+  attn_implementation: sdpa
+
+dataloader:
+  _target_: torchdata.stateful_dataloader.StatefulDataLoader
+  num_workers: 1
+  persistent_workers: true
+  pin_memory: true
+  drop_last: true
+
+validation_dataset: none
+validation_dataloader: none
+
+optimizer:
+  _target_: torch.optim.AdamW
+  betas: [0.9, 0.95]
+  eps: 1e-8
+  lr: 5.0e-6
+  weight_decay: 0.1
+
+lr_scheduler:
+  lr_decay_style: constant
+
+clip_grad_norm:
+  max_norm: 1.0
+
+wandb:
+  enable: false
+  project: huiyingl_workspace
+  entity: Nemo-automodel
+  name: glm5_3_flash_medpix_packed2k_ep72_cp2_100steps
+  group: glm5_3_flash_medpix_packed2k_cp_parity_100steps
+  tags: [glm5.3-flash, medpix, vlm, packed2k, ep72, cp2, parity]
+  dir: logs/glm5_3_flash_medpix_packed2k_ep72_cp2_100steps/wandb
+
+ci:
+  recipe_owner: HuiyingLi
+  nodes: 9
+  time: "01:00:00"
+  max_steps: 5

nemo_automodel/_transformers/registry.py

@@ -86,6 +86,13 @@
             "GlmMoeDsaForCausalLM",
             ("nemo_automodel.components.models.glm_moe_dsa.model", "GlmMoeDsaForCausalLM"),
         ),
+        (
+            "Glm5NextForConditionalGeneration",
+            (
+                "nemo_automodel.components.models.glm5_next.model",
+                "Glm5NextForConditionalGeneration",
+            ),
+        ),
         (
             "Gemma4ForConditionalGeneration",
             ("nemo_automodel.components.models.gemma4_moe.model", "Gemma4ForConditionalGeneration"),
@@ -326,6 +333,7 @@
     "bailing_moe": ("nemo_automodel.components.models.ling_v2.config", "BailingMoeV2Config"),
     "deepseek_v4": ("nemo_automodel.components.models.deepseek_v4.config", "DeepseekV4Config"),
     "glm_moe_dsa": ("nemo_automodel.components.models.glm_moe_dsa.config", "GlmMoeDsaConfig"),
+    "glm5_next": ("nemo_automodel.components.models.glm5_next.config", "Glm5NextConfig"),
     "hy_v3": ("nemo_automodel.components.models.hy_v3.config", "HYV3Config"),
     "inkling_mm_model": ("nemo_automodel.components.models.inkling.configuration", "InklingConfig"),
     "kimi_k2": ("nemo_automodel.components.models.kimi_k2.config", "KimiK2Config"),