[AutoRound] Support w8a8 scheme in auto-round and add example

examples/autoround/README.md

@@ -16,7 +16,7 @@ pip install -e .
 
 ## Quickstart
 
-The example includes an end-to-end script for applying the AutoRound quantization algorithm.
+The [example](./quantization_w4a16/llama3_example.py) includes an end-to-end script for applying the AutoRound quantization algorithm.
 
 ```bash
 python3 llama3_example.py
@@ -134,7 +134,7 @@ We can see the resulting scores look good!
 > Note: quantized model accuracy may vary slightly due to nondeterminism.
 
 ### Known Issues
-Currently, `llm-compressor` supports applying AutoRound only on the `wNa16` quantization schemes. Support for additional schemes is planned. You can follow progress in the [RFC](https://github.com/vllm-project/llm-compressor/issues/1968).
+Currently, `llm-compressor` supports applying AutoRound only on the `wNa16` and `w8a8` quantization schemes. Support for additional schemes is planned. You can follow progress in the [RFC](https://github.com/vllm-project/llm-compressor/issues/1968).
 
 ### Questions or Feature Request?
 

examples/autoround/quantization_w8a8_fp8/llama4_dynamic_quant_example.py

@@ -0,0 +1,61 @@
+from auto_round.calib_dataset import get_dataset
+from transformers import AutoTokenizer, Llama4ForConditionalGeneration, AutoProcessor
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.autoround import AutoRoundModifier
+from llmcompressor.utils import dispatch_for_generation
+
+# Select model and load it.
+model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+model = Llama4ForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+processor = AutoProcessor.from_pretrained(model_id)
+
+# Select calibration dataset.
+NUM_CALIBRATION_SAMPLES = 128
+MAX_SEQUENCE_LENGTH = 2048
+# Get aligned calibration dataset.
+
+ds = get_dataset(
+    tokenizer=tokenizer,
+    seqlen=MAX_SEQUENCE_LENGTH,
+    nsamples=NUM_CALIBRATION_SAMPLES,
+)
+
+
+# Configure the quantization algorithm to run.
+scheme = "FP8_DYNAMIC"
+recipe = AutoRoundModifier(
+    targets="Linear",
+    scheme=scheme,
+    ignore=["re:.*lm_head", "re:.*router", "re:.*self_attn.*", "re:.*shared_expert.*" , "re:multi_modal_projector.*", "re:vision_model"],
+    iters=0,
+)
+
+
+# Apply algorithms.
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    # disable shuffling to get slightly better mmlu score
+    shuffle_calibration_samples=False,
+)
+
+# Confirm generations of the quantized model look sane.
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+dispatch_for_generation(model)
+sample = tokenizer("Hello my name is", return_tensors="pt")
+sample = {key: value.to(model.device) for key, value in sample.items()}
+output = model.generate(**sample, max_new_tokens=1)
+print(tokenizer.decode(output[0]))
+print("==========================================\n\n")
+
+# Save to disk compressed.
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W8A8-Dynamic-AutoRound"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)
+processor.save_pretrained(SAVE_DIR)

examples/autoround/quantization_w8a8_fp8/llama4_static_quant_example.py

@@ -0,0 +1,61 @@
+from auto_round.calib_dataset import get_dataset
+from transformers import AutoTokenizer, Llama4ForConditionalGeneration, AutoProcessor
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.autoround import AutoRoundModifier
+from llmcompressor.utils import dispatch_for_generation
+
+# Select model and load it.
+model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+model = Llama4ForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+processor = AutoProcessor.from_pretrained(model_id)
+
+# Select calibration dataset.
+NUM_CALIBRATION_SAMPLES = 128
+MAX_SEQUENCE_LENGTH = 2048
+# Get aligned calibration dataset.
+
+ds = get_dataset(
+    tokenizer=tokenizer,
+    seqlen=MAX_SEQUENCE_LENGTH,
+    nsamples=NUM_CALIBRATION_SAMPLES,
+)
+
+
+# Configure the quantization algorithm to run.
+scheme = "FP8"
+recipe = AutoRoundModifier(
+    targets="Linear",
+    scheme=scheme,
+    ignore=["re:.*lm_head", "re:.*router", "re:.*self_attn.*", "re:.*shared_expert.*" , "re:multi_modal_projector.*", "re:vision_model"],
+    iters=0,
+)
+
+
+# Apply algorithms.
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    # disable shuffling to get slightly better mmlu score
+    shuffle_calibration_samples=False,
+)
+
+# Confirm generations of the quantized model look sane.
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+dispatch_for_generation(model)
+sample = tokenizer("Hello my name is", return_tensors="pt")
+sample = {key: value.to(model.device) for key, value in sample.items()}
+output = model.generate(**sample, max_new_tokens=1)
+print(tokenizer.decode(output[0]))
+print("==========================================\n\n")
+
+# Save to disk compressed.
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W8A8-Dynamic-AutoRound"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)
+processor.save_pretrained(SAVE_DIR)

src/llmcompressor/modifiers/autoround/base.py

@@ -3,6 +3,8 @@
 import torch
 from auto_round import AutoRound
 from auto_round.schemes import QuantizationScheme as ARQuantizationScheme
+from auto_round.wrapper import WrapperWALayer
+
 from compressed_tensors.quantization import (
     QuantizationScheme,
     QuantizationStrategy,
@@ -145,8 +147,7 @@ def start_calibration(self, model: torch.nn.Module):
             untie_word_embeddings(model)
 
         for _, module in match_named_modules(model, self.targets, self.ignore):
-            # Note: No need to register observers for auto-round
-            self._calibration_hooks |= self._initialize_hooks(module)
+            # skip register observers for auto-round
             apply_calibration_status(module)
 
         model.apply(enable_quantization)  # quantize at the same time as calibrate
@@ -242,16 +243,38 @@ def apply_autoround(self, state, subgraph):
                 auto_offload=False,
             )
             self._q_input = q_input
+
+            # auto-round will return WrapperWALayer if activation is quantized
+            for name, module in decoding_layer.named_modules():
+                if isinstance(module, WrapperWALayer):
+                    parent, child = name.rsplit(".", maxsplit=1)
+                    parent = decoding_layer.get_submodule(parent)
+                    setattr(parent, child, module.orig_layer)
+
             # Update offload parameters and remove temporary attributes
-            for _, module in decoding_layer.named_modules():
-                if hasattr(module, "weight_scale") and hasattr(
-                    module, "weight_zero_point"
+            for name, module in decoding_layer.named_modules():
+                if (
+                    hasattr(module, "weight_scale")
+                    and hasattr(module, "weight_zero_point")
+                    and hasattr(module, "scale")
                 ):
                     # Note: The model's weight is already q-dq in-place by auto-round.
                     weight_scale = module.scale
                     del module.scale
                     # TODO: update zero_point after supporting asymmetric quantization
                     update_offload_parameter(module, "weight_scale", weight_scale)
+
+                if (
+                    hasattr(module, "act_scale")
+                    and hasattr(module, "input_scale")
+                ):
+                    act_scale = module.act_scale
+                    assert act_scale.numel() == module.input_scale.numel(), "Activation scale size mismatch"
+                    del module.act_scale
+
+                    # activation scale shape maybe different
+                    update_offload_parameter(module, "input_scale", act_scale.reshape(module.input_scale.shape))
+
         decoding_layer.eval()
 
     def post_autoround_cleanup(self):
@@ -314,23 +337,65 @@ def _mapping_config_to_autoround(self):
             ), f"Expected QuantizationScheme, got {type(scheme)}"
             quant_scheme = scheme
         weight_args = quant_scheme.weights
-        assert weight_args.strategy == QuantizationStrategy.GROUP, (
-            "Only group-wise quantization is supported in AutoRoundModifier for now, "
-            f"got {weight_args.strategy}"
-        )
-        assert quant_scheme.input_activations is None, (
-            "Input activation quantization is not supported in AutoRoundModifier, "
-            f"got {quant_scheme.input_activations}"
-        )
+        activation_args = quant_scheme.input_activations
         assert quant_scheme.output_activations is None, (
             "Output activation quantization is not supported in AutoRoundModifier, "
             f"got {quant_scheme.output_activations}"
         )
+        group_size = weight_args.group_size
+        data_type = weight_args.type
+        if group_size is None:
+            if weight_args.strategy == QuantizationStrategy.CHANNEL:
+                group_size = -1
+            elif weight_args.strategy == QuantizationStrategy.TENSOR:
+                group_size = 0
+            else:
+                raise ValueError(
+                    "AutoRoundModifier only supports channel-wise and tensor-wise weight quantization"
+                )
+
+        if data_type == "float":
+            data_type = "fp"
+
+        if activation_args is None:
+            act_bits = 16
+            act_group_size = None
+            act_symmetric = None
+            act_bits = None
+            act_dynamic = None
+            act_data_type = None
+        else:
+            act_dynamic = activation_args.dynamic
+            act_group_size = activation_args.group_size
+            act_symmetric = activation_args.symmetric
+            act_bits = activation_args.num_bits
+
+            # activation is quantized dynamically, don't need to collect scale in auto-round
+            if act_dynamic:
+                act_bits = 16
+
+            act_data_type = activation_args.type
+            assert activation_args.strategy != QuantizationStrategy.GROUP, (
+                "Input activation group-wise quantization is not supported in AutoRoundModifier"
+            )
+            if act_group_size is None:
+                if activation_args.strategy in [QuantizationStrategy.CHANNEL, QuantizationStrategy.TOKEN]:
+                    group_size = -1
+                if activation_args.strategy == QuantizationStrategy.TENSOR:
+                    group_size = 0
+
+            if act_data_type == "float":
+                act_data_type = "fp"
+
         ar_quant_scheme = ARQuantizationScheme(
             bits=weight_args.num_bits,
             sym=weight_args.symmetric,
-            group_size=weight_args.group_size,
-            data_type=weight_args.type,
-            act_bits=16,
+            group_size=group_size,
+            data_type=data_type,
+            act_bits=act_bits,
+            act_group_size=act_group_size,
+            act_sym=act_symmetric,
+            act_dynamic=act_dynamic,
+            act_data_type=act_data_type,
         )
         return ar_quant_scheme

tests/llmcompressor/transformers/autoround/test_autoround_oneshot.py

@@ -39,6 +39,29 @@
     },
 )
 
+rtn_recipe_modifier = AutoRoundModifier(
+    ignore=["lm_head"],
+    iters=0,
+    config_groups={
+        "group_0": QuantizationScheme(
+            targets=["Linear"],
+            weights=QuantizationArgs(num_bits=8, strategy="group", group_size=128),
+            input_activations=QuantizationArgs(num_bits=8, type="float", strategy="tensor", dynamic=True),
+        )
+    },
+)
+
+act_quant_recipe_modifier = AutoRoundModifier(
+    ignore=["lm_head"],
+    iters=0,
+    config_groups={
+        "group_0": QuantizationScheme(
+            targets=["Linear"],
+            weights=QuantizationArgs(num_bits=8, type="float", strategy="channel"),
+            input_activations=QuantizationArgs(num_bits=8, type="float", strategy="tensor"),
+        )
+    },
+)
 
 @requires_gpu(1)
 @pytest.mark.parametrize(
@@ -93,4 +116,57 @@ def test_oneshot_application(recipe, tmp_path):
 
     # Check lm-head is not quantized
     not_targetted = model_loaded.lm_head
+
+@requires_gpu(1)
+@pytest.mark.parametrize(
+    "recipe",
+    [
+        rtn_recipe_modifier,
+        act_quant_recipe_modifier
+    ],
+)
+def test_rtn_oneshot(recipe, tmp_path):
+    output = tmp_path / "oneshot_output"
+    model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
+    tokenizer = AutoTokenizer.from_pretrained(model)
+    dataset = get_dataset(
+        tokenizer=tokenizer,
+        seqlen=1024,
+        nsamples=32,
+    )
+
+    device = "cuda:0" if torch.cuda.is_available() else "cpu"
+
+    oneshot(
+        model=model,
+        dataset=dataset,
+        output_dir=output,
+        recipe=recipe,
+    )
+    model_loaded = AutoModelForCausalLM.from_pretrained(output, device_map=device)
+
+    quantization_config = model_loaded.config.quantization_config.quantization_config
+    assert quantization_config is not None
+
+    # check config is set properly
+    assert "lm_head" in quantization_config.ignore
+    assert len(quantization_config.config_groups) == 1
+    quant_scheme = quantization_config.config_groups["group_0"]
+    assert isinstance(quant_scheme, QuantizationScheme)
+
+    weight_args = quantization_config.config_groups["group_0"].weights
+    act_args = quantization_config.config_groups["group_0"].input_activations
+    assert isinstance(weight_args, QuantizationArgs)
+    assert weight_args.num_bits == recipe.config_groups["group_0"].weights.num_bits
+    assert weight_args.strategy == recipe.config_groups["group_0"].weights.strategy
+    if act_args is not None:
+        assert act_args.num_bits == recipe.config_groups["group_0"].input_activations.num_bits
+        assert act_args.strategy== recipe.config_groups["group_0"].input_activations.strategy
+
+    # Check a specific layer is quantized
+    targetted_linear_layer = model_loaded.model.layers[2].self_attn.q_proj
+    assert hasattr(targetted_linear_layer, "quantization_scheme")
+
+    # Check lm-head is not quantized
+    not_targetted = model_loaded.lm_head
     assert not hasattr(not_targetted, "quantization_scheme")