Add Qwen3 0.6B, 1.7B, and 4B

README.md

@@ -51,7 +51,7 @@ To get started you can:
 
 - Visit the [Step by Step Tutorial](https://pytorch.org/executorch/stable/getting-started.html) to get things running locally and deploy a model to a device
 - Use this [Colab Notebook](https://colab.research.google.com/drive/1qpxrXC3YdJQzly3mRg-4ayYiOjC6rue3?usp=sharing) to start playing around right away
-- Jump straight into LLM use cases by following specific instructions for [Llama](examples/models/llama/README.md) and [Llava](examples/models/llava/README.md)
+- Jump straight into LLM use cases by following specific instructions for popular open-source models such as [Llama](examples/models/llama/README.md), [Qwen 3](examples/models/qwen3/README.md), [Phi-4-mini](examples/models/phi_4_mini/README.md), and [Llava](examples/models/llava/README.md)
 
 ## Feedback and Engagement
 

examples/models/llama/attention.py

@@ -178,6 +178,7 @@ def __init__(self, args: ModelArgs, layer_id: int, rope: Rope):
         self.dim = args.dim
         self.attention_qkv_bias = args.attention_qkv_bias
         self.use_qk_norm = args.use_qk_norm
+        self.qk_norm_before_rope = args.qk_norm_before_rope
 
         if self.use_qk_norm:
             q_norm_dim = self.head_dim
@@ -243,14 +244,18 @@ def forward(
         k = k.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
         v = v.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
 
+        if self.use_qk_norm and self.qk_norm_before_rope:
+            q = self.q_norm_fn(q)
+            k = self.k_norm_fn(k)
+
         # RoPE relative positional embeddings
         q, k = self.rope.forward(q, k, freqs_cos, freqs_sin)
 
         q = q.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
         k = k.transpose(1, 2)
         v = v.transpose(1, 2)
 
-        if self.use_qk_norm:
+        if self.use_qk_norm and not self.qk_norm_before_rope:
             q = self.q_norm_fn(q)
             k = self.k_norm_fn(k)
 

examples/models/llama/export_llama_lib.py

@@ -100,6 +100,9 @@
     "llama3_2",
     "static_llama",
     "qwen2_5",
+    "qwen3-0_6b",
+    "qwen3-1_7b",
+    "qwen3-4b",
     "phi_4_mini",
     "smollm2",
 ]
@@ -108,6 +111,9 @@
     "qwen2_5": "Qwen/Qwen2.5-1.5B",
     "phi_4_mini": "microsoft/Phi-4-mini-instruct",
     "smollm2": "HuggingFaceTB/SmolLM-135M",
+    "qwen3-0_6b": "Qwen/Qwen3-0.6B",
+    "qwen3-1_7b": "Qwen/Qwen3-1.7B",
+    "qwen3-4b": "Qwen/Qwen3-4B",
 }
 
 
@@ -544,6 +550,10 @@ def export_llama(args) -> str:
             from executorch.examples.models.qwen2_5 import (  # pyre-ignore[21]
                 convert_weights,
             )
+        elif args.model.startswith("qwen3"):
+            from executorch.examples.models.qwen3 import (  # pyre-ignore[21]
+                convert_weights,
+            )
         elif args.model == "phi_4_mini":
             from executorch.examples.models.phi_4_mini import (  # pyre-ignore[21]
                 convert_weights,

examples/models/llama/model_args.py

@@ -38,6 +38,7 @@ class ModelArgs:
     apply_embedding: bool = True  # Use embedding inside the transformer
     apply_output: bool = True  # Use output layer (unembedding) inside the transformer
     use_qk_norm: bool = False  # apply normalization to q and k in the attention
+    qk_norm_before_rope: bool = False  # when to apply qk norm
     use_hf_rope: bool = False  # Use HuggingFace's RoPE implementation
     partial_rotary_factor: float = 1.0
     rope_theta: Optional[float] = (

examples/models/qwen3/0_6b_config.json

@@ -0,0 +1,17 @@
+{
+  "dim": 1024,
+  "ffn_dim_multiplier": 1,
+  "hidden_dim": 3072,
+  "n_heads": 16,
+  "head_dim": 128,
+  "n_kv_heads": 8,
+  "n_layers": 28,
+  "norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "use_scaled_rope": false,
+  "vocab_size": 151936,
+  "use_hf_rope": true,
+  "attention_qkv_bias": false,
+  "use_qk_norm": true,
+  "qk_norm_before_rope": true
+}

examples/models/qwen3/1_7b_config.json

@@ -0,0 +1,17 @@
+{
+  "dim": 2048,
+  "ffn_dim_multiplier": 1,
+  "hidden_dim": 6144,
+  "n_heads": 16,
+  "head_dim": 128,
+  "n_kv_heads": 8,
+  "n_layers": 28,
+  "norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "use_scaled_rope": false,
+  "vocab_size": 151936,
+  "use_hf_rope": true,
+  "attention_qkv_bias": false,
+  "use_qk_norm": true,
+  "qk_norm_before_rope": true
+}

examples/models/qwen3/4b_config.json

@@ -0,0 +1,17 @@
+{
+  "dim": 2560,
+  "ffn_dim_multiplier": 1,
+  "hidden_dim": 9728,
+  "n_heads": 32,
+  "head_dim": 128,
+  "n_kv_heads": 8,
+  "n_layers": 36,
+  "norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "use_scaled_rope": false,
+  "vocab_size": 151936,
+  "use_hf_rope": true,
+  "attention_qkv_bias": false,
+  "use_qk_norm": true,
+  "qk_norm_before_repo": true
+}

examples/models/qwen3/README.md

@@ -0,0 +1,85 @@
+## Summary
+Qwen 3 is the latest iteration of the Qwen series of large language models (LLMs) developed by Alibaba. Edge-sized Qwen3 model variations (0.6B, 1.7B, and 4B) are currently supported .
+
+## Instructions
+
+Qwen 3 uses the same example code as our optimized Llama model, while the checkpoint, model params, and tokenizer are different. Please see the [Llama README page](../llama/README.md) for details.
+
+All commands for exporting and running Llama on various backends should also be applicable to Qwen 3, by swapping the following args:
+```
+--model [qwen3-0.6b,qwen3-1_7b,qwen3-4b]
+--params [examples/models/qwen3/0_6b_config.json,examples/models/qwen3/1_7b_config.json,examples/models/qwen3/4b_config.json]
+```
+
+### Example export
+Here is a basic example for exporting Qwen 3, although please refer to the Llama README's [Step 2: Prepare model](../llama/README.md#step-2-prepare-model) for more advanced usage.
+
+Export 0.6b to XNNPack, quantized with 8da4w:
+```
+python -m examples.models.llama.export_llama \
+  --model qwen3-0_6b \
+  --params examples/models/qwen3/0_6b_config.json \
+  -kv \
+  --use_sdpa_with_kv_cache \
+  -d fp32 \
+  -X \
+  --xnnpack-extended-ops \
+  -qmode 8da4w
+  --output_name="qwen3-0_6b.pte" \
+  --verbose
+```
+
+Export 1.7b to XNNPack, quantized with 8da4w:
+```
+python -m examples.models.llama.export_llama \
+  --model qwen3-1_7b \
+  --params examples/models/qwen3/1_7b_config.json \
+  -kv \
+  --use_sdpa_with_kv_cache \
+  -d fp32 \
+  -X \
+  --xnnpack-extended-ops \
+  -qmode 8da4w
+  --output_name="qwen3-1_7b.pte" \
+  --verbose
+```
+
+Export 4b to XNNPack, quantized with 8da4w:
+```
+python -m examples.models.llama.export_llama \
+  --model qwen3-4b \
+  --params examples/models/qwen3/4b_config.json \
+  -kv \
+  --use_sdpa_with_kv_cache \
+  -d fp32 \
+  -X \
+  --xnnpack-extended-ops \
+  -qmode 8da4w
+  --output_name="qwen3-4b.pte" \
+  --verbose
+```
+
+### Example run
+With ExecuTorch pybindings:
+```
+python -m examples.models.llama.runner.native
+  --model qwen3-0_6b \
+  --pte qwen3-0_6b.pte \
+  --tokenizer ~/.cache/huggingface/hub/models--Qwen--Qwen3-0.6B/snapshots/a9c98e602b9d36d2a2f7ba1eb0f5f31e4e8e5143/tokenizer.json \
+  --tokenizer_config ~/.cache/huggingface/hub/models--Qwen--Qwen3-0.6B/snapshots/a9c98e602b9d36d2a2f7ba1eb0f5f31e4e8e5143/tokenizer_config.json \
+  --prompt "Who is the president of the US?" \
+  --params examples/models/qwen3/0_6b_config.json \
+  --max_len 128 \
+  -kv \
+  --temperature 0.6
+```
+
+With ExecuTorch's sample c++ runner (see the Llama README's [Step 3: Run on your computer to validate](../llama/README.md#step-3-run-on-your-computer-to-validate) to build the runner):
+```
+cmake-out/examples/models/llama/llama_main
+  --model_path qwen3-0_6b.pte
+  --tokenizer_path ~/.cache/huggingface/hub/models--Qwen--Qwen3-0.6B/snapshots/a9c98e602b9d36d2a2f7ba1eb0f5f31e4e8e5143/tokenizer.json
+  --prompt="Who is the president of the US?"
+```
+
+To run the model on an example iOS or Android app, see the Llama README's [Step 5: Build Mobile apps](../llama/README.md#step-5-build-mobile-apps) section.

examples/models/qwen3/__init__.py

@@ -0,0 +1,16 @@
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from executorch.examples.models.llama.model import Llama2Model
+from executorch.examples.models.qwen3.convert_weights import convert_weights
+
+
+class Qwen3Model(Llama2Model):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+
+__all__ = [
+    "Qwen3Model",
+    "convert_weights",
+]

examples/models/qwen3/convert_weights.py

@@ -0,0 +1,115 @@
+import argparse
+
+import json
+import os
+from typing import Dict
+
+import torch
+from safetensors.torch import load_file
+
+from torchtune.models.convert_weights import get_mapped_key
+
+# Standard _FROM_META weight mapping of Meta weights to TorchTune + additional bias weight mappings.
+_QWEN_3_FROM_META = {
+    "tok_embeddings.weight": "model.embed_tokens.weight",
+    "norm.weight": "model.norm.weight",
+    "layers.{}.attention.wk.weight": "model.layers.{}.self_attn.k_proj.weight",
+    "layers.{}.attention.k_norm_fn.weight": "model.layers.{}.self_attn.k_norm.weight",
+    "layers.{}.attention.wq.weight": "model.layers.{}.self_attn.q_proj.weight",
+    "layers.{}.attention.q_norm_fn.weight": "model.layers.{}.self_attn.q_norm.weight",
+    "layers.{}.attention.wv.weight": "model.layers.{}.self_attn.v_proj.weight",
+    "layers.{}.attention.wo.weight": "model.layers.{}.self_attn.o_proj.weight",
+    "layers.{}.attention_norm.weight": "model.layers.{}.input_layernorm.weight",
+    "layers.{}.ffn_norm.weight": "model.layers.{}.post_attention_layernorm.weight",
+    # Note: gate_proj and up_proj are reversed, usually w1 is the up_proj,
+    # w2 is the gate_proj, and activation is applied on the up_proj, but since
+    # Qwen3 applies activation on the gate_proj, we just swap the gate_proj
+    # and up_proj in the checkpoint itself as a hack.
+    "layers.{}.feed_forward.w1.weight": "model.layers.{}.mlp.gate_proj.weight",
+    "layers.{}.feed_forward.w2.weight": "model.layers.{}.mlp.down_proj.weight",
+    "layers.{}.feed_forward.w3.weight": "model.layers.{}.mlp.up_proj.weight",
+}
+
+
+def qwen_3_tune_to_meta(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
+    """
+    Convert a state dict from torchtune's format to Meta's format. This function
+    doesn't handle any sharding or splitting of state dicts. It follows the
+    state_dict IN -> state_dict OUT pattern.
+
+    Args:
+        state_dict (Dict[str, torch.Tensor]): State dict in torchtune's format.
+
+    Returns:
+        Dict[str, torch.Tensor]: State dict in Meta's format.
+    """
+    converted_state_dict = {}
+    inverted_mapping_dict = {v: k for k, v in _QWEN_3_FROM_META.items()}
+
+    for key, value in state_dict.items():
+        # Tied embeddings for 0.6b and 4b models.
+        if key == "lm_head.weight":
+            continue
+        new_key = get_mapped_key(key, inverted_mapping_dict)
+        converted_state_dict[new_key] = value
+
+    converted_state_dict["output.weight"] = converted_state_dict[
+        "tok_embeddings.weight"
+    ]
+
+    return converted_state_dict
+
+
+def load_checkpoint(input_dir: str) -> Dict:
+    index_path = os.path.join(input_dir, "model.safetensors.index.json")
+    if os.path.exists(index_path):
+        # Sharded checkpoint.
+        with open(index_path, "r") as f:
+            index = json.load(f)
+        weight_map = index["weight_map"]
+        checkpoint_shards = sorted(set(weight_map.values()))
+
+        # Load all the shards into memory
+        shard_to_weights = {}
+        for shard in checkpoint_shards:
+            shard_to_weights[shard] = load_file(os.path.join(input_dir, shard))
+
+        # Merge tensors into consolidated state dict.
+        merged_state_dict = {}
+        for weight_name, shard in weight_map.items():
+            tensor = shard_to_weights[shard][weight_name]
+            merged_state_dict[weight_name] = tensor
+        return merged_state_dict
+    else:
+        # Single checkpoint.
+        state_dict = load_file(os.path.join(input_dir, "model.safetensors"))
+        return state_dict
+
+
+def convert_weights(input_dir: str, output_file: str) -> None:
+    print("Loading checkpoint...")
+    sd = load_checkpoint(input_dir)
+    print("Converting checkpoint...")
+    sd = qwen_3_tune_to_meta(sd)
+    print("Saving checkpoint...")
+    torch.save(sd, output_file)
+    print("Done.")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert Qwen3 weights to Meta format."
+    )
+    parser.add_argument(
+        "input_dir",
+        type=str,
+        help="Path to directory containing checkpoint files",
+    )
+    parser.add_argument("output", type=str, help="Path to the output checkpoint")
+
+    args = parser.parse_args()
+    convert_weights(args.input_dir, args.output)
+
+
+if __name__ == "__main__":
+    main()