From 2b79f143756bf10a2d43e5ab6d413374b633d2ce Mon Sep 17 00:00:00 2001
From: 44670 <44670@users.noreply.github.com>
Date: Tue, 15 Jul 2025 17:53:41 +0800
Subject: [PATCH] support loading qwen3 gguf (#38645)

* support loading qwen3 gguf

* Add qwen3 into GGUF_TO_FAST_CONVERTERS for tokenizer conversion

* Add testcase

* Fix formatting
---
 src/transformers/integrations/ggml.py | 13 +++++++++++++
 tests/quantization/ggml/test_ggml.py  | 18 ++++++++++++++++++
 2 files changed, 31 insertions(+)

diff --git a/src/transformers/integrations/ggml.py b/src/transformers/integrations/ggml.py
index 51bdc88608..17f86e1667 100644
--- a/src/transformers/integrations/ggml.py
+++ b/src/transformers/integrations/ggml.py
@@ -90,6 +90,18 @@ GGUF_CONFIG_MAPPING = {
         "expert_count": "num_experts",
         "expert_used_count": "num_experts_per_tok",
     },
+    "qwen3": {
+        "context_length": "max_position_embeddings",
+        "block_count": "num_hidden_layers",
+        "feed_forward_length": "intermediate_size",
+        "embedding_length": "hidden_size",
+        "rope.dimension_count": None,
+        "rope.freq_base": "rope_theta",
+        "attention.head_count": "num_attention_heads",
+        "attention.head_count_kv": "num_key_value_heads",
+        "attention.layer_norm_rms_epsilon": "rms_norm_eps",
+        "vocab_size": "vocab_size",
+    },
     "falcon": {
         "context_length": "max_position_embeddings",
         "block_count": "num_hidden_layers",
@@ -676,6 +688,7 @@ GGUF_TO_FAST_CONVERTERS = {
     "llama": GGUFLlamaConverter,
     "qwen2": GGUFQwen2Converter,
     "qwen2_moe": GGUFQwen2Converter,
+    "qwen3": GGUFQwen2Converter,
     "phi3": GGUFPhi3Converter,
     "bloom": GGUFGPTConverter,
     "falcon": GGUFGPTConverter,
diff --git a/tests/quantization/ggml/test_ggml.py b/tests/quantization/ggml/test_ggml.py
index 6a0321f065..d2ed7f7a74 100644
--- a/tests/quantization/ggml/test_ggml.py
+++ b/tests/quantization/ggml/test_ggml.py
@@ -301,6 +301,7 @@ class GgufModelTests(unittest.TestCase):
     gemma3_qat_model_id = "google/gemma-3-1b-it-qat-q4_0-gguf"
     gemma3_text_model_id = "unsloth/gemma-3-1b-it-GGUF"
     gemma3_vision_model_id = "unsloth/gemma-3-4b-it-GGUF"
+    qwen3_model_id = "Qwen/Qwen3-0.6B-GGUF"
 
     q4_0_phi3_model_id = "Phi-3-mini-4k-instruct-q4.gguf"
     q4_0_mistral_model_id = "mistral-7b-instruct-v0.2.Q4_0.gguf"
@@ -333,6 +334,7 @@ class GgufModelTests(unittest.TestCase):
     q4_0_gemma3_qat_model_id = "gemma-3-1b-it-q4_0.gguf"
     bf16_gemma3_text_model_id = "gemma-3-1b-it-BF16.gguf"
     bf16_gemma3_vision_model_id = "gemma-3-4b-it-BF16.gguf"
+    q8_0_qwen3_model_id = "Qwen3-0.6B-Q8_0.gguf"
 
     example_text = "Hello"
 
@@ -955,3 +957,19 @@ class GgufModelTests(unittest.TestCase):
                 torch.testing.assert_close(original_params, converted_state_dict[layer_name])
             else:
                 raise ValueError(f"Layer {layer_name} is not presented in GGUF model")
+
+    @require_read_token
+    @unittest.skipUnless(is_gguf_available("0.16.0"), "test requires gguf version >= 0.16.0")
+    def test_qwen3_q8_0(self):
+        tokenizer = AutoTokenizer.from_pretrained(self.qwen3_model_id, gguf_file=self.q8_0_qwen3_model_id)
+        model = AutoModelForCausalLM.from_pretrained(
+            self.qwen3_model_id,
+            gguf_file=self.q8_0_qwen3_model_id,
+            torch_dtype=torch.float16,
+        )
+
+        text = tokenizer(self.example_text, return_tensors="pt")["input_ids"]
+        out = model.generate(text, max_new_tokens=10)
+
+        EXPECTED_TEXT = "HelloED\nI need to find the value of the"
+        self.assertEqual(tokenizer.decode(out[0], skip_special_tokens=True), EXPECTED_TEXT)