[GGUF] support Qwen3 architecture

[TianmengChen]

Detail:

1. Load attn_q_norm attn_k_norm weight for Qwen3 architecture.
2. Add rms_norm in splite_head after q k head for Qwen3 architecture.

Validation Scope:
Qwen3-0.6B-f16 (CPU/GPU)
Qwen3-0.6B-Q8_0 (CPU*)
Qwen3-4B-Q4_K_M (CPU/GPU)

*Q8_0 has accuracy issue on GPU: CVS-166108

[Copilot]

Pull Request Overview

This PR adds support for the Qwen3 architecture by updating model configuration, weight loading, and head-splitting logic. Key changes include:

• Adding Qwen3 to the list of supported architectures in gguf_modeling.cpp.
• Updating head_size configuration to use the key_length metadata if available and loading attn_q_norm/attn_k_norm weights in gguf.cpp.
• Introducing a new split_heads helper function in building_blocks.cpp that applies RMS normalization for Qwen3 and updating multi_head_attention to use it.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/cpp/src/gguf_utils/gguf_modeling.cpp	Expanded architecture support by including Qwen3 in model creation logic.
src/cpp/src/gguf_utils/gguf.cpp	Adjusted head_size calculation and added loading for Qwen3-specific normalization weights.
src/cpp/src/gguf_utils/building_blocks.cpp	Added a new split_heads helper function with RMS norm handling and updated multi_head_attention to use it.

Comments suppressed due to low confidence (1)

src/cpp/src/gguf_utils/building_blocks.cpp:357

• The key used for v-split normalization ("self_attn.v_norm") does not appear to be loaded anywhere, unlike the Qwen3 q_norm and k_norm weights. Please confirm if the absence of a v_norm weight is intentional or if the weight loading logic should be updated accordingly.

auto v_split = split_heads(value, num_heads_kv, head_dim, rms_norm_eps, key_name + ".self_attn.v_norm", consts);

[sammysun0711]

build_jenkins

[sammysun0711]

Please add a test case for Qwen3: e.g. Qwen/Qwen3-0.6B-GGUF

• Add gguf_model_id and gguf_filename in get_gguf_model_list:

  openvino.genai/tests/python_tests/data/models.py

  Lines 79 to 82 in 8988da2

  	{
  	"gguf_model_id": "Qwen/Qwen2.5-0.5B-Instruct-GGUF",
  	"gguf_filename": "qwen2.5-0.5b-instruct-q4_0.gguf"
  	},

• Make sure test passed:

  openvino.genai/tests/python_tests/test_llm_pipeline.py

  Line 850 in 8988da2

  def test_full_gguf_pipeline(pipeline_type, model_ids):

[sammysun0711]

It seems that Qwen3 GGUF was not supported by transformer yet: ggml.py
Hence the test failed here: https://github.com/openvinotoolkit/openvino.genai/actions/runs/15337247120/job/43157146883?pr=2273
GGUF model with architecture qwen3 is not supported yet.
@rkazants, can we skip the test first after qwen3 GGUF support by transformers?

[rkazants]

It seems that Qwen3 GGUF was not supported by transformer yet: ggml.py Hence the test failed here: https://github.com/openvinotoolkit/openvino.genai/actions/runs/15337247120/job/43157146883?pr=2273 GGUF model with architecture qwen3 is not supported yet. @rkazants, can we skip the test first after qwen3 GGUF support by transformers?

please add a separate test without comparation with HF results. Just let us compare with some hard-coded expected result for input prompt. And add a comment that this is temporal testing solution until transformer starts to support qwen3 in GGUF format.

[sammysun0711]

build_jenkins

[sammysun0711]

build_jenkins

[sammysun0711]

It seems that Qwen3 GGUF was not supported by transformer yet: ggml.py Hence the test failed here: https://github.com/openvinotoolkit/openvino.genai/actions/runs/15337247120/job/43157146883?pr=2273 GGUF model with architecture qwen3 is not supported yet. @rkazants, can we skip the test first after qwen3 GGUF support by transformers?

please add a separate test without comparation with HF results. Just let us compare with some hard-coded expected result for input prompt. And add a comment that this is temporal testing solution until transformer starts to support qwen3 in GGUF format.

Qwen3 GGUF related test added:

openvino.genai/tests/python_tests/test_llm_pipeline.py

Line 887 in 82b53d4

def test_full_gguf_qwen3_pipeline(pipeline_type, model_ids):

CI test run & passed: https://github.com/openvinotoolkit/openvino.genai/actions/runs/15546210687/job/43773618720?pr=2273#step:8:151

[sammysun0711]

build_jenkins