[WIP] Qwen3 MoE support

recipes/configs/qwen3/30B_A3B_full.yaml

@@ -0,0 +1,120 @@
+# Config for multi-device full finetuning in full_finetune_distributed.py
+# using a Qwen3 30B A3B model
+#
+# This config assumes that you've run the following command before launching:
+#   tune download Qwen/Qwen3-30B-A3B --output-dir /tmp/Qwen3-30B-A3B
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config qwen3/30B_A3B_full
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config qwen3/30B_A3B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+
+output_dir: /tmp/qwen3_30B_A3B/full # /tmp may be deleted by your system. Change it to your preference.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen3.qwen3_tokenizer
+  path: /tmp/Qwen3-30B-A3B/vocab.json
+  merges_file: /tmp/Qwen3-30B-A3B/merges.txt
+  max_seq_len: null
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+  packed: False  # True increases speed
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen3.qwen3_moe_30b_a3b_instruct
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen3-30B-A3B
+  checkpoint_files: [
+    model-00001-of-00016.safetensors,
+    model-00002-of-00016.safetensors,
+    model-00003-of-00016.safetensors,
+    model-00004-of-00016.safetensors,
+    model-00005-of-00016.safetensors,
+    model-00006-of-00016.safetensors,
+    model-00007-of-00016.safetensors,
+    model-00008-of-00016.safetensors,
+    model-00009-of-00016.safetensors,
+    model-00010-of-00016.safetensors,
+    model-00011-of-00016.safetensors,
+    model-00012-of-00016.safetensors,
+    model-00013-of-00016.safetensors,
+    model-00014-of-00016.safetensors,
+    model-00015-of-00016.safetensors,
+    model-00016-of-00016.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN3_MOE
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  lr: 5e-6
+loss:
+  _component_: torchtune.modules.loss.LinearCrossEntropyLoss
+max_steps_per_epoch: null
+gradient_accumulation_steps: 8  # Use to increase effective batch size
+clip_grad_norm: null
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+#custom_sharded_layers: ['tok_embeddings', 'output']  # Layers to shard separately (useful for large vocab size models). Lower Memory, but lower speed.
+#fsdp_cpu_offload: True
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 1
+log_peak_memory_stats: True
+log_level: INFO  # DEBUG, WARN, etc.
+
+
+# Profiler (disabled)
+profiler:
+  _component_: torchtune.training.setup_torch_profiler
+  enabled: False
+
+  #Output directory of trace artifacts
+  output_dir: ${output_dir}/profiling_outputs
+
+  #`torch.profiler.ProfilerActivity` types to trace
+  cpu: True
+  cuda: True
+
+  #trace options passed to `torch.profiler.profile`
+  profile_memory: False
+  with_stack: False
+  record_shapes: True
+  with_flops: False
+
+  # `torch.profiler.schedule` options:
+  # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
+  wait_steps: 5
+  warmup_steps: 3
+  active_steps: 2
+  num_cycles: 1

recipes/lora_finetune_distributed.py

@@ -272,11 +272,11 @@ def setup(self, cfg: DictConfig) -> None:
             self._metric_logger.log_config(cfg)
 
         if (
-            cfg.checkpointer.model_type == "LLAMA4"
+            cfg.checkpointer.model_type in ("LLAMA4", "QWEN3_MOE")
             and self._save_adapter_weights_only is False
         ):
             raise ValueError(
-                "For Llama4 training, you should set save_adapter_weights_only to True."
+                f"For {cfg.checkpointer.model_type} training, you should set save_adapter_weights_only to True."
             )
 
         checkpoint_dict = self._checkpoint_client.load_base_checkpoint()

torchtune/models/qwen3/__init__.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from ._component_builders import lora_qwen3_moe, qwen3_moe
 from ._model_builders import (
     lora_qwen3_0_6b_base,
     lora_qwen3_0_6b_instruct,
@@ -16,6 +17,9 @@
     lora_qwen3_4b_instruct,
     lora_qwen3_8b_base,
     lora_qwen3_8b_instruct,
+    lora_qwen3_moe_235b_a22b,
+    lora_qwen3_moe_30b_a3b_base,
+    lora_qwen3_moe_30b_a3b_instruct,
     qwen3_0_6b_base,
     qwen3_0_6b_instruct,
     qwen3_14b_base,
@@ -27,10 +31,18 @@
     qwen3_4b_instruct,
     qwen3_8b_base,
     qwen3_8b_instruct,
+    qwen3_moe_235b_a22b,
+    qwen3_moe_30b_a3b_base,
+    qwen3_moe_30b_a3b_instruct,
     qwen3_tokenizer,
 )
 
+from ._parallelism import tp_plan
+
 __all__ = [
+    "tp_plan",
+    "qwen3_moe",
+    "lora_qwen3_moe",
     "lora_qwen3_0_6b_base",
     "lora_qwen3_0_6b_instruct",
     "lora_qwen3_1_7b_base",
@@ -42,6 +54,9 @@
     "lora_qwen3_14b_base",
     "lora_qwen3_14b_instruct",
     "lora_qwen3_32b",
+    "lora_qwen3_moe_30b_a3b_base",
+    "lora_qwen3_moe_30b_a3b_instruct",
+    "lora_qwen3_moe_235b_a22b",
     "qwen3_0_6b_base",
     "qwen3_0_6b_instruct",
     "qwen3_1_7b_base",
@@ -53,5 +68,8 @@
     "qwen3_14b_base",
     "qwen3_14b_instruct",
     "qwen3_32b",
+    "qwen3_moe_30b_a3b_base",
+    "qwen3_moe_30b_a3b_instruct",
+    "qwen3_moe_235b_a22b",
     "qwen3_tokenizer",
 ]