cleanup

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

examples/transform/llama3_example.py

@@ -2,13 +2,11 @@
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from llmcompressor import oneshot
-from llmcompressor.modifiers.quantization import GPTQModifier, QuantizationModifier
+from llmcompressor.modifiers.quantization import QuantizationModifier
 from llmcompressor.modifiers.transform import SpinQuantModifier
 from llmcompressor.utils import dispatch_for_generation
 
 # Select model and load it.
-# MODEL_ID = "meta-llama/Llama-3.2-1B-Instruct"
-# MODEL_ID = "meta-llama/Llama-3.2-3B-Instruct" # TODO hidden size 3072 causes failure when creating hadamard
 MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
 
 model = AutoModelForCausalLM.from_pretrained(
@@ -57,36 +55,32 @@ def tokenize(sample):
 ds = ds.map(tokenize, remove_columns=ds.column_names)
 
 # Configure the quantization algorithm to run.
+#   * apply spinquant transforms to model in order to make quantization easier
 #   * quantize the weights to 4 bit with GPTQ with a group size 128
 recipe = [
-    # TODO preset_config="QUIP_ONLINE" outputs gibberish
-    # preset_config="QUIP" output sensible, but cannot load saved
-    #  checkpoint or run evals (~4hrs to run)
-    SpinQuantModifier(rotations=["R1", "R2"]),
-    # QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
+    SpinQuantModifier(rotations=["R1", "R2"], transform_type="random-hadamard"),
+    QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
 ]
 
 # Apply algorithms.
 oneshot(
     model=model,
     recipe=recipe,
-    # dataset=ds,
-    pipeline="datafree",
-    # max_seq_length=MAX_SEQUENCE_LENGTH,
-    # num_calibration_samples=NUM_CALIBRATION_SAMPLES,
-    log_dir=None,
+    dataset=ds,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
 )
 
-# # Confirm generations of the quantized model look sane.
+# Confirm generations of the quantized model look sane.
 print("\n\n")
 print("========== SAMPLE GENERATION ==============")
 dispatch_for_generation(model)
 input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
 output = model.generate(input_ids, max_new_tokens=100)
 print(tokenizer.decode(output[0]))
-# print("==========================================\n\n")
+print("==========================================\n\n")
 
-# # Save to disk compressed.
-# SAVE_DIR = MODEL_ID.split("/")[1] + "-transform-quant-w4a16"
-# model.save_pretrained(SAVE_DIR, save_compressed=True)
-# tokenizer.save_pretrained(SAVE_DIR)
+# Save to disk compressed.
+SAVE_DIR = MODEL_ID.split("/")[1] + "-transformed-w4a16"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

src/llmcompressor/entrypoints/oneshot.py

@@ -125,8 +125,7 @@ def __init__(
         self.output_dir = output_dir
 
         # initialize the model and processor
-        # TODO Remove Comment before merge, this is just needed for DummyModel
-        # pre_process(model_args)
+        pre_process(model_args)
 
         # Set instance attributes
         self.model = self.model_args.model

examples/transform/spinquant_dummy.py renamed to tests/llmcompressor/modifiers/transform/test_dummy_model.py

@@ -14,9 +14,6 @@
 num_embeddings = 12
 
 
-# TODO remove file before merging
-
-
 class DummySelfAttn(torch.nn.Module):
     def __init__(self, hidden_dim, intermediate_dim):
         super().__init__()
@@ -75,37 +72,36 @@ def forward(self, input_ids):
         return self.lm_head(x)
 
 
-model = DummyModel(num_embeddings, hidden_dim, intermediate_dim, up_dim)
-
-# TODO Uncomment this to see norm diff > 1e-6
-# This is due to issue Kyle spotted in https://arxiv.org/pdf/2405.16406 Page 5 Footnote 2
-# Will have to fuse layernorms with subsequent layers so that input_layernorm.weight is equal to torch.ones() (this apparently makes it rotation invariant)
-# https://github.com/facebookresearch/SpinQuant/blob/8f47aa3f00e8662caf1a484153920a07e5281c3a/utils/fuse_norm_utils.py#L39
-# update_parameter_data(
-#     model.input_layernorm,
-#     torch.rand(model.input_layernorm.weight.shape),
-#     "weight",
-# )
-
-input_ids = torch.IntTensor([1, 2, 3, 4, 5])
-orig_output = model(input_ids)
-
-recipe = [
-    # NOTE: preset_config="QUIP" output sensible, but cannot load saved
-    #  checkpoint or run evals (~4hrs to run)
-    SpinQuantModifier(rotations=["R1", "R2"]),
-    # QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
-]
-
-oneshot(
-    model=model,
-    recipe=recipe,
-    pipeline="datafree",
-    log_dir=None,
-)
-
-# # Confirm generations of the quantized model look the same
-transformed_output = model(input_ids)
-
-print(f"Norm Diff {(orig_output-transformed_output).norm()}")
-print(f"Norm {orig_output.norm()}, {transformed_output.norm()}")
+def test_dummy_model():
+    model = DummyModel(num_embeddings, hidden_dim, intermediate_dim, up_dim)
+
+    # TODO Uncomment this to see norm diff > 1e-6
+    # This is due to issue Kyle spotted in https://arxiv.org/pdf/2405.16406 Page 5 Footnote 2
+    # Will have to fuse layernorms with subsequent layers so that input_layernorm.weight is equal to torch.ones() (this apparently makes it rotation invariant)
+    # https://github.com/facebookresearch/SpinQuant/blob/8f47aa3f00e8662caf1a484153920a07e5281c3a/utils/fuse_norm_utils.py#L39
+    # update_parameter_data(
+    #     model.input_layernorm,
+    #     torch.rand(model.input_layernorm.weight.shape),
+    #     "weight",
+    # )
+
+    input_ids = torch.IntTensor([1, 2, 3, 4, 5])
+    orig_output = model(input_ids)
+
+    recipe = [
+        SpinQuantModifier(rotations=["R1", "R2"]),
+    ]
+
+    # TODO: work around preprocessing?
+    oneshot(
+        model=model,
+        recipe=recipe,
+        pipeline="datafree",
+        log_dir=None,
+    )
+
+    # # Confirm generations of the quantized model look the same
+    transformed_output = model(input_ids)
+
+    print(f"Norm Diff {(orig_output-transformed_output).norm()}")
+    print(f"Norm {orig_output.norm()}, {transformed_output.norm()}")