[Qwen2_5_vl] - Onboarding Qwen2_5_vl model in QEfficient (#560)

Signed-off-by: Mohit Soni <[email protected]>
Co-authored-by: Hem Agnihotri <[email protected]>

Author: mohiso22

QEfficient/transformers/models/modeling_auto.py

@@ -1313,9 +1313,14 @@ def kv_offload_generate(
         vision_end = perf_counter()
 
         lang_inputs = {k: v for k, v in inputs.items() if k not in vision_inputs}
-        lang_inputs["position_ids"] = np.where(
-            lang_inputs.pop("attention_mask"), np.arange(padded_len), -1
-        )  # Need to use -1 as position_ids for invalid tokens
+
+        if "position_ids" in inputs:
+            lang_inputs["position_ids"] = inputs["position_ids"]
+            lang_inputs.pop("attention_mask")
+        else:
+            lang_inputs["position_ids"] = np.where(
+                lang_inputs.pop("attention_mask"), np.arange(padded_len), -1
+            )  # Need to use -1 as position_ids for invalid tokens
 
         not_mllama = hasattr(self.model.config, "model_type") and self.model.config.model_type != "mllama"
         if not_mllama:
@@ -1336,7 +1341,7 @@ def kv_offload_generate(
         for i in range(num_chunks):
             chunk_inputs["input_ids"] = lang_inputs["input_ids"][:, i * prefill_seq_len : (i + 1) * prefill_seq_len]
             chunk_inputs["position_ids"] = lang_inputs["position_ids"][
-                :, i * prefill_seq_len : (i + 1) * prefill_seq_len
+                ..., i * prefill_seq_len : (i + 1) * prefill_seq_len
             ]
             outputs = lang_session.run(chunk_inputs)
             chunk_inputs["image_idx"] = outputs["image_idx_output"]
@@ -1353,7 +1358,7 @@ def kv_offload_generate(
 
         # Get first token
         lang_inputs["input_ids"] = outputs["logits"].argmax(2)
-        lang_inputs["position_ids"] = input_len.numpy()
+        lang_inputs["position_ids"] = np.max(lang_inputs["position_ids"], axis=-1, keepdims=True) + 1
         if "cross_attention_mask" in lang_inputs:
             bs, _, num_images, img_tiles = lang_inputs["cross_attention_mask"].shape
             lang_inputs["cross_attention_mask"] = torch.ones((bs, 1, num_images, img_tiles), dtype=torch.int64).numpy()

QEfficient/transformers/models/pytorch_transforms.py

@@ -152,6 +152,18 @@
     Qwen2Model,
     Qwen2RMSNorm,
 )
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    Qwen2_5_VisionTransformerPretrainedModel,
+    Qwen2_5_VLAttention,
+    Qwen2_5_VLDecoderLayer,
+    Qwen2_5_VLForConditionalGeneration,
+    Qwen2_5_VLModel,
+    Qwen2_5_VLTextModel,
+    Qwen2_5_VLVisionAttention,
+)
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    Qwen2RMSNorm as Qwen2_5RMSNorm,
+)
 from transformers.models.qwen3.modeling_qwen3 import (
     Qwen3Attention,
     Qwen3DecoderLayer,
@@ -356,6 +368,15 @@
     QEffQwen2ForCausalLM,
     QEffQwen2Model,
 )
+from QEfficient.transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    QEffQwen2_5_VisionTransformerPretrainedModel,
+    QEffQwen2_5_VLAttention,
+    QEffQwen2_5_VLDecoderLayer,
+    QEffQwen2_5_VLModel,
+    QEffQwen2_5_VLTextModel,
+    QEffQwen2_5_VLVisionAttention,
+    QEffQwen_2_5_vl_ForConditionalGeneration,
+)
 from QEfficient.transformers.models.qwen3.modeling_qwen3 import (
     QEffQwen3Attention,
     QEffQwen3DecoderLayer,
@@ -404,6 +425,7 @@ class CustomOpsTransform(ModuleMappingTransform):
         Phi3RMSNorm: CustomRMSNormAIC,
         Qwen2RMSNorm: CustomRMSNormAIC,
         Qwen3RMSNorm: CustomRMSNormAIC,
+        Qwen2_5RMSNorm: CustomRMSNormAIC,
         MllamaTextRMSNorm: CustomRMSNormAIC,
         GraniteRMSNorm: CustomRMSNormAIC,
         PixtralRMSNorm: CustomRMSNormAIC,
@@ -544,6 +566,14 @@ class KVCacheTransform(ModuleMappingTransform):
         Qwen3DecoderLayer: QEffQwen3DecoderLayer,
         Qwen3Model: QEffQwen3Model,
         Qwen3ForCausalLM: QEffQwen3ForCausalLM,
+        # Qwen2.5 VL
+        Qwen2_5_VLForConditionalGeneration: QEffQwen_2_5_vl_ForConditionalGeneration,
+        Qwen2_5_VLModel: QEffQwen2_5_VLModel,
+        Qwen2_5_VLAttention: QEffQwen2_5_VLAttention,
+        Qwen2_5_VLDecoderLayer: QEffQwen2_5_VLDecoderLayer,
+        Qwen2_5_VisionTransformerPretrainedModel: QEffQwen2_5_VisionTransformerPretrainedModel,
+        Qwen2_5_VLVisionAttention: QEffQwen2_5_VLVisionAttention,
+        Qwen2_5_VLTextModel: QEffQwen2_5_VLTextModel,
         # Starcoder2
         Starcoder2Attention: QEffStarcoder2Attention,
         Starcoder2DecoderLayer: QEFFStarcoder2DecoderLayer,

QEfficient/transformers/models/qwen2_5_vl/__init__.py

@@ -0,0 +1,6 @@
+# -----------------------------------------------------------------------------
+#
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+# SPDX-License-Identifier: BSD-3-Clause
+#
+# -----------------------------------------------------------------------------