Feat: Onbaord PlamoForCausalLM Architecture

QEfficient/base/pytorch_transforms.py

@@ -107,6 +107,9 @@ def apply(cls, model: nn.Module) -> Tuple[nn.Module, bool]:
             ):
                 for orig_method_name, mapped_method in repl_method_map.items():
                     setattr(module, orig_method_name, MethodType(mapped_method, module))
+                    # Handling the __init__ calls in the models
+                    if hasattr(module, "__qeff_init__"):
+                        module.__qeff_init__()
                     transformed = True
 
         return model, transformed

QEfficient/transformers/models/modeling_auto.py

@@ -1304,6 +1304,7 @@ class QEFFAutoModelForCausalLM(QEFFBaseModel):
         FP8DeQuantLinearToLinearTransform,
         CustomOpsTransform,
         KVCacheTransform,
+        KVCacheModuleMethodMapperTransform,
     ]
     _onnx_transforms = [FP16ClipTransform, SplitTensorsTransform]
 

QEfficient/transformers/models/plamo/__init__.py

@@ -0,0 +1,6 @@
+# -----------------------------------------------------------------------------
+#
+# Copyright (c) 2025 Qualcomm Innovation Center, Inc. All rights reserved.
+# SPDX-License-Identifier: BSD-3-Clause
+#
+# -----------------------------------------------------------------------------

QEfficient/transformers/models/pytorch_transforms.py

@@ -245,6 +245,13 @@
     QEffPhi3ForCausalLM,
     QEffPhi3Model,
 )
+from QEfficient.transformers.models.plamo.modeling_plamo import (
+    QEffPlamoAttention,
+    QEffPlamoDecoder,
+    QEffPlamoDecoderLayer,
+    QEffPlamoForCausalLM,
+    QEffPlamoModel,
+)
 from QEfficient.transformers.models.qwen2.modeling_qwen2 import (
     QEffQwen2Attention,
     QEffQwen2DecoderLayer,
@@ -485,5 +492,10 @@ class KVCacheModuleMethodMapperTransform(ModuleMethodMapperTransform):
             "get_qeff_language_decoder": QEffInternVLModel.get_qeff_language_decoder,
         },
         "InternVisionEmbeddings": {"forward": QEffInternVisionEmbeddings.forward},
+        "PlamoForCausalLM": {"forward": QEffPlamoForCausalLM.forward},
+        "PlamoModel": {"forward": QEffPlamoModel.forward},
+        "PlamoDecoder": {"forward": QEffPlamoDecoder.forward},
+        "PlamoDecoderLayer": {"forward": QEffPlamoDecoderLayer.forward},
+        "Attention": {"forward": QEffPlamoAttention.forward},
     }
     _match_class_replace_method = {}

README.md

@@ -20,6 +20,7 @@
 - [04/2025] [Granite 3.0 and 3.1 Language MOE Models] (https://huggingface.co/ibm-granite/granite-3.0-1b-a400m-base)
 - [09/2024] [AWQ](https://arxiv.org/abs/2306.00978)/[GPTQ](https://arxiv.org/abs/2210.17323) 4-bit quantized models are supported <br>
 - [09/2024] Now we support [PEFT](https://huggingface.co/docs/peft/index) models
+- [04/2025] Added support for [PLaMo] (https://huggingface.co/pfnet/plamo-13b-instruct)
 - [01/2025] Added support for [Ibm-Granite] (https://huggingface.co/ibm-granite/granite-3.1-8b-instruct)
 - [01/2025] Added support for [Ibm-Granite-Guardian] (https://huggingface.co/ibm-granite/granite-guardian-3.1-8b)
 - [09/2024] Added support for [Gemma-2-Family](https://huggingface.co/collections/google/gemma-2-release-667d6600fd5220e7b967f315)<br>

tests/transformers/models/test_causal_lm_models.py

@@ -44,6 +44,7 @@
     "neuralmagic/Qwen2-0.5B-Instruct-FP8",  # fp8 quant method, static, with lm head ignored
     "ibm-granite/granite-3.1-2b-instruct",
     "ibm-granite/granite-guardian-3.1-2b",
+    "pfnet/plamo-13b-instruct",
 ]
 
 test_models_qnn = [