Added quantization for evaluation script (pytorch#11822)

Summary:
Pull Request resolved: pytorch#11822

Added quantization to evaluation script. Quantization causes deterioriation in accuracy

On wikitext task:
| Model Name | max_seq_len | ptq | word_perplexity
|----------|----------|----------|-----------|
| Llama 3.2-1B Instruct  | 128   | 16a4w |  5821003.055178451 |
| Llama 3.2-1B Instruct  | 128   | 16a4w_block |  5396240.078572427 |
| Llama 3.2-1B Instruct  | 128   | 8a8w |  533154.970440251 |

Reviewed By: cccclai

Differential Revision: D76837572

Author: rohansjoshi

examples/qualcomm/oss_scripts/llama/TARGETS

@@ -49,6 +49,9 @@ python_binary(
     name = "eval_llama_qnn",
     srcs = ["eval_llama_qnn.py"],
     main_function = "executorch.examples.qualcomm.oss_scripts.llama.eval_llama_qnn.main",
+    preload_deps = [
+        "//executorch/extension/llm/custom_ops:model_sharding_py",
+    ],
     deps = [
         ":llama_lib",
         "//executorch/examples/models/llama:eval_library",

examples/qualcomm/oss_scripts/llama/eval_llama_qnn.py

@@ -4,9 +4,14 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import sys
 import argparse
 import copy
 import json
+import torch
+from functools import partial
+
+from lm_eval.evaluator import simple_evaluate
 
 from typing import List, Optional, Tuple
 
@@ -26,32 +31,53 @@
 
 from pytorch_tokenizers import get_tokenizer
 
+from executorch.examples.qualcomm.oss_scripts.llama.llama import calibrate
+
+from executorch.examples.qualcomm.utils import make_quantizer
+
+from executorch.examples.models.llama.source_transformation.quantize import (
+    get_quant_embedding_transform,
+)
+
+from torchao.quantization.pt2e import MinMaxObserver
+from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
+
+
+from executorch.backends.qualcomm.quantizer.quantizer import QuantDtype
+from executorch.backends.qualcomm.utils.utils import convert_linear_to_conv2d
+from executorch.backends.qualcomm.quantizer.custom_annotation import (
+    annotate_linear_16a8w_in_affine_layer,
+    annotate_matmul_16a8w,
+)
+
+
+import logging
+sys.setrecursionlimit(4096)
+FORMAT = "[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s"
+logging.basicConfig(level=logging.INFO, format=FORMAT)
+logging.getLogger().setLevel(logging.INFO)
+
 
 class WrappedLlamaModel(nn.Module):
-    def __init__(self, model, use_kv_cache=False, max_seq_len=512, device="cuda"):
+    def __init__(self, model, atten_mask, use_kv_cache=False, max_seq_len=512, device="cuda"):
         super(WrappedLlamaModel, self).__init__()
         self.model = model
         self.max_seq_len = max_seq_len
         self.use_kv_cache = use_kv_cache
         self.device = device
+        self.atten_mask = atten_mask
 
     def forward(
         self,
         tokens: torch.Tensor,
-        input_pos: Optional[torch.Tensor] = None,
         *args,
     ) -> Tuple[torch.Tensor, List[torch.Tensor], List[torch.Tensor]]:
         # Pad input if necessary, since LlamaModel requires static shape
         if tokens.shape[1] != self.max_seq_len:
             tokens = torch.nn.functional.pad(
-                tokens, (self.max_seq_len - tokens.shape[1], 0)
+                tokens, (0, self.max_seq_len - tokens.shape[1])
             )
-        atten_mask = (
-            self.model.get_example_inputs(self.use_kv_cache)[1]
-            .to(device=self.device)
-            .to(dtype=torch.bfloat16)
-        )
-        return self.model.forward(tokens, atten_mask, input_pos, *args)
+        return self.model.forward(tokens, self.atten_mask)
 
 
 def gen_eval_wrapper(model_name, args):
@@ -119,14 +145,73 @@ def permute(w, heads):
             layer.feed_forward.prepare_feedfoward_conv()
 
     model.to(dtype=torch.bfloat16)
-    model.to(args.device)
+    model.to(device=args.device)
+
+    tokens, atten_mask = model.get_example_inputs(use_kv_cache=False)
+    tokens = tokens.to(device=args.device)
+    atten_mask = atten_mask.to(device=args.device)
+    atten_mask = atten_mask.to(dtype=torch.bfloat16)
+    inputs = (tokens, atten_mask)
+
+    if args.embedding_quantize:
+        model = get_quant_embedding_transform(
+            embedding_quantize=args.embedding_quantize
+        )(model)
+
+    model = convert_linear_to_conv2d(model)
 
-    wrapped_model = WrappedLlamaModel(
-        model, args.use_kv_cache, args.max_seq_length, args.device
+    if args.ptq:
+        quant_dtype = getattr(QuantDtype, f"use_{args.ptq}")
+
+        custom_annotations = (annotate_matmul_16a8w,)
+        if args.llama_model == "stories110m":
+            custom_annotations = custom_annotations + (
+                annotate_linear_16a8w_in_affine_layer,
+            )
+        quantizer = make_quantizer(
+            quant_dtype=quant_dtype,
+            per_channel_conv=True,
+            per_channel_linear=True,
+            act_observer=MinMaxObserver,
+        )
+        quantizer.add_custom_quant_annotations(custom_annotations)
+
+        model.has_quant_io = True
+
+        with torch.no_grad():
+            model = torch.export.export(
+                model, inputs, strict=True
+            ).module()
+            if quant_dtype == QuantDtype.use_16a4w_block:
+                conv_nodes = [
+                    n for n in model.graph.nodes if "conv" in n.name
+                ]
+                block_size_map = {n.name: (1, 64, 1, 1) for n in conv_nodes}
+                quantizer.set_block_size_map(block_size_map)
+
+            model = prepare_pt2e(model, quantizer)
+
+        logging.info("Quantizing the model...")
+
+        calibrate(
+            inputs,
+            'Once upon a time',
+            model,
+            tokenizer=tokenizer,
+            ar_len=args.prefill_ar_len,
+            max_seq_len=args.max_seq_len,
+            kv_updater=None,
+            use_i64_token=use_i64_token,
+        )
+
+        model = convert_pt2e(model)
+
+    model = WrappedLlamaModel(
+        model, atten_mask, args.use_kv_cache, args.max_seq_length, args.device
     )
 
     return GraphModuleEvalWrapper(
-        model=wrapped_model,
+        model=model,
         tokenizer=tokenizer,
         max_seq_length=args.calibration_seq_length,
         use_kv_cache=args.use_kv_cache,
@@ -167,6 +252,7 @@ def main() -> None:
     modelname = "llama2"
     parser = build_args_parser()
     args = parser.parse_args()
+    args.llama_model = "llama3_2"
     # Overrides this arg, because evaluation requires full logits.
     args.generate_full_logits = True
 
@@ -177,7 +263,15 @@ def main() -> None:
     args.use_kv_cache = False
     args.prefill_ar_len = args.max_seq_length
 
+    # To do fewer samples for faster evaluation
+    args.limit = 0.1
+    # args.samples = {'wikitext': list(range(1))}
+
     args.device = "cuda" if torch.cuda.is_available() else "cpu"
+    torch.set_default_device(args.device)
+
+    args.ptq = '8a8w'
+
 
     eval_llama(modelname, args)