Enable Vulkan 4-bit weight only quantization in export_llama (#6235)

Summary:
Pull Request resolved: #6235

As title.
ghstack-source-id: 248349849
exported-using-ghexport

Reviewed By: jorgep31415

Differential Revision: D64406456

fbshipit-source-id: 20890a7391f821b9b58063bef305909d34d48a18

Author: SS-JIA

backends/vulkan/_passes/TARGETS

@@ -62,6 +62,7 @@ runtime.python_library(
     ],
     visibility = [
         "//executorch/backends/...",
+        "//executorch/examples/...",
     ],
     deps = [
         ":int4_weight_only_quantizer",

examples/models/llama2/TARGETS

@@ -103,6 +103,7 @@ runtime.python_library(
     deps = [
         "//ai_codesign/gen_ai/fast_hadamard_transform:fast_hadamard_transform",
         "//caffe2:torch",
+        "//executorch/backends/vulkan/_passes:vulkan_passes",
         "//executorch/examples/models:model_base",
         "//executorch/examples/models:models",
         "//executorch/extension/llm/custom_ops:custom_ops_aot_py",

examples/models/llama2/export_llama_lib.py

@@ -157,7 +157,7 @@ def build_args_parser() -> argparse.ArgumentParser:
         "--quantization_mode",
         type=str,
         default=None,
-        choices=["int8", "8da4w", "8da4w-gptq"],
+        choices=["int8", "8da4w", "8da4w-gptq", "vulkan_4w"],
         help="type of quantization",
     )
 

examples/models/llama2/source_transformation/quantize.py

@@ -12,6 +12,8 @@
 import torch.nn as nn
 import torch.nn.functional as F
 
+from executorch.backends.vulkan._passes import VkInt4WeightOnlyQuantizer
+
 from executorch.extension.llm.export.builder import DType
 
 from sentencepiece import SentencePieceProcessor
@@ -31,7 +33,7 @@
     fsLinear = nn.Linear
 
 
-def quantize(
+def quantize(  # noqa C901
     model: torch.nn.Module,
     qmode: str,
     activation_dtype: Optional[DType],
@@ -131,6 +133,9 @@ def quantize(
         )
         model = gptq_quantizer.quantize(model, inputs)
         return model
+    elif qmode == "vulkan_4w":
+        model = VkInt4WeightOnlyQuantizer().quantize(model)
+        return model
     else:
         raise Exception(f"Unrecognized quantize mode: {qmode}")
 

extension/llm/export/partitioner_lib.py

@@ -37,9 +37,6 @@ def get_vulkan_partitioner(
     assert (
         dtype_override == "fp32" or dtype_override is None
     ), "Vulkan backend does not support non fp32 dtypes at the moment"
-    assert (
-        quantization_mode is None
-    ), "Vulkan backend does not support quantization at the moment"
     from executorch.backends.vulkan.partitioner.vulkan_partitioner import (
         VulkanPartitioner,
     )