Fix hash for VLM's language decoder to include qaic_config

Signed-off-by: quic-xiyushi <xiyushi@qti.qualcomm.com>

Author: quic-xiyushi

QEfficient/transformers/models/modeling_auto.py

@@ -8,7 +8,7 @@
 import warnings
 from pathlib import Path
 from time import perf_counter
-from typing import Dict, List, Optional, Union
+from typing import List, Optional, Union
 
 import numpy as np
 import torch
@@ -752,19 +752,22 @@ class QEffCausalLMForTextImageToTextModel(QEFFBaseModel):
     ]
     _onnx_transforms = [FP16ClipTransform, SplitTensorsTransform]
 
-    def __init__(self, model, **kwargs):
+    def __init__(self, model, qaic_config: Optional[dict] = None, **kwargs):
         """
         Initializes the language decoder component for multimodal models.
 
         Parameters
         ----------
         model : nn.Module
             The full HuggingFace multimodal model from which the language decoder is extracted.
+        qaic_config : dict, optional
+            A dictionary for QAIC-specific configurations.
         **kwargs :
             Additional keyword arguments passed to the base class constructor.
         """
-        super().__init__(model, **kwargs)
+        super().__init__(model, qaic_config=qaic_config, **kwargs)
         self.model = model.get_qeff_language_decoder()
+        self.model.qaic_config = qaic_config
         self.hash_params["qeff_auto_class"] = self.__class__.__name__
 
     def export(
@@ -936,9 +939,8 @@ def __init__(
         self.comp_ctx_lengths_prefill, self.comp_ctx_lengths_decode = process_ccl_specializations(qaic_config)
 
         self.vision_model = QEffVisionEncoderForTextImageToTextModel(model, **kwargs)
-        self.lang_model = QEffCausalLMForTextImageToTextModel(model, **kwargs)
+        self.lang_model = QEffCausalLMForTextImageToTextModel(model, qaic_config=qaic_config, **kwargs)
         self.continuous_batching = continuous_batching
-        self.lang_model.model.qaic_config = qaic_config
         self.input_shapes, self.output_names = None, None
         # ---Sampling---
         # Note: SamplerTransform should be applied after all other transforms
@@ -2286,7 +2288,6 @@ def from_pretrained(
             logger.warning("Updating low_cpu_mem_usage=False")
 
         kwargs.update({"attn_implementation": "eager", "low_cpu_mem_usage": False})
-
         model = cls._hf_auto_class.from_pretrained(pretrained_model_name_or_path, **kwargs)
         return cls(
             model,
@@ -2365,6 +2366,8 @@ def __init__(
             - **return_pdfs** (bool): If True, returns probability distributions along with sampled tokens.
               For Speculative Decoding Target Language Models, this is always True.
             - **max_top_k_ids** (int): Maximum number of top K tokens (<= vocab size) to consider during sampling.
+            - **include_guided_decoding** (bool): If True, enables guided token-level filtering
+              during decoding. Only works when include_sampler=True.
         **kwargs :
             Additional keyword arguments passed to the base class constructor.
 
@@ -2467,6 +2470,8 @@ def from_pretrained(
               and ``return_pdfs=False`` for regular model.
             - **max_top_k_ids** (int): Maximum number of top K tokens (<= vocab size) to consider during sampling.
               The values provided in ``top_ks`` tensor must be less than this maximum limit.
+            - **include_guided_decoding** (bool): If True, enables guided token-level filtering
+              during decoding. Only works when include_sampler=True.
 
         *args :
             Positional arguments passed directly to `cls._hf_auto_class.from_pretrained`.