Update MoE examples (#192)

* Update MoE examples

* Add top-level link

* Fix deepseek_moe_w8a8_int8.py

* Add deepseek_moe_w8a8_fp8.py

* Quality

* Quality

Author: mgoin

README.md

@@ -38,6 +38,7 @@ Applying quantization with `llmcompressor`:
 * [Activation quantization to `int8`](examples/quantization_w8a8_int8)
 * [Activation quantization to `fp8`](examples/quantization_w8a8_fp8)
 * [Weight only quantization to `int4`](examples/quantization_w4a16)
+* [Quantizing MoE LLMs](examples/quantizing_moe)
 
 ### User Guides
 Deep dives into advanced usage of `llmcompressor`:

examples/quantizing_moe/README.md

@@ -1,6 +1,6 @@
-# Quantizing TinyMixtral-4x248M-MoE Model with FP8
+# Quantizing Mixtral-8x7B-Instruct-v0.1 Model with FP8
 
-This directory contains an example script for quantizing the `TinyMixtral-4x248M-MoE` model using the FP8 quantization scheme.
+This directory contains an example script for quantizing the `Mixtral-8x7B-Instruct-v0.1` model using the static per-tensor FP8 quantization scheme.
 
 ## Installation
 
@@ -17,7 +17,7 @@ pip install -e .
 The provided example script demonstrates an end-to-end process for applying the quantization algorithm:
 
 ```bash
-python3 mixtral_moe_fp8.py
+python3 mixtral_moe_w8a8_fp8.py
 ```
 
 ## Creating a Quantized MoE Model
@@ -27,7 +27,7 @@ This example leverages `llm-compressor` and `compressed-tensors` to create an FP
 You can follow the detailed steps below or simply run the example script with:
 
 ```bash
-python examples/quantizing_moe/mixtral_moe_fp8.py
+python mixtral_moe_w8a8_fp8.py
 ```
 
 ### Step 1: Select a Model, Dataset, and Recipe
@@ -36,12 +36,12 @@ In this step, you'll choose a baseline model for quantization, a dataset for cal
 
 - **Models**: Can be referenced from a local directory or retrieved from the Hugging Face Hub.
 - **Datasets**: Can also be from a local directory or the Hugging Face Hub.
-- **Recipes**: These are YAML files or Python modifier objects that describe how a model should be optimized during or after training. In this example, we use a `GPTQModifier` object with the scheme set to `FP8`.
+- **Recipes**: These are YAML files or Python modifier objects that describe how a model should be optimized during or after training. In this example, we use a `QuantizationModifier` object with the scheme set to `FP8`.
 
 ```python
-from llmcompressor.modifiers.quantization.gptq import GPTQModifier
+from llmcompressor.modifiers.quantization import QuantizationModifier
 
-recipe = GPTQModifier(scheme="FP8", targets="Linear", ignore=["lm_head", "re:.*block_sparse_moe.gate"], sequential_update=True)
+recipe = QuantizationModifier(scheme="FP8", targets="Linear", ignore=["lm_head", "re:.*block_sparse_moe.gate"])
 ```
 
 NOTE: `.*block_sparse_moe.gate` layers do not quantize well, hence they are ignored!
@@ -69,9 +69,11 @@ oneshot(
 
 ### Custom Quantization
 
+NOTE: Only per-tensor quantization is supported in vLLM as of now (`vllm==0.6.1`)
+
 The repository supports multiple quantization techniques configured via a recipe. Supported strategies include `tensor`, `group`, and `channel` quantization.
 
-In the above example, FP8 channel-wise quantization is used as specified by the `FP8` scheme. For other preset schemes, refer to the [quantization schemes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/quant_scheme.py) in the `Compressed-Tensors` library.
+In the above example, FP8 per-tensor quantization is used as specified by the `FP8` scheme. For other preset schemes, refer to the [quantization schemes](https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/quant_scheme.py) in the `compressed-tensors` library.
 
 A custom scheme can also be specified using `config_groups`:
 
@@ -89,7 +91,7 @@ config_groups = {
                         "num_bits": 8,
                         "type": "int",
                         "symmetric": true,
-                        "strategy": "tensor",
+                        "strategy": "group",
                         "group_size": 128, 
                     }
                }

examples/quantizing_moe/deepseek_moe_w8a8_fp8.py

@@ -0,0 +1,83 @@
+from datasets import load_dataset
+from transformers import AutoTokenizer
+
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
+
+# select a Mixture of Experts model for quantization
+MODEL_ID = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct"
+
+model = SparseAutoModelForCausalLM.from_pretrained(
+    MODEL_ID, device_map="auto", torch_dtype="auto", trust_remote_code=True
+)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+# Select calibration dataset.
+# its recommended to use more calibration samples for MoE models so each expert is hit
+DATASET_ID = "HuggingFaceH4/ultrachat_200k"
+DATASET_SPLIT = "train_sft"
+NUM_CALIBRATION_SAMPLES = 2048
+MAX_SEQUENCE_LENGTH = 2048
+
+
+# Load dataset and preprocess.
+ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
+ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
+
+
+def preprocess(example):
+    return {
+        "text": tokenizer.apply_chat_template(
+            example["messages"],
+            tokenize=False,
+        )
+    }
+
+
+ds = ds.map(preprocess)
+
+
+# Tokenize inputs.
+def tokenize(sample):
+    return tokenizer(
+        sample["text"],
+        padding=False,
+        max_length=MAX_SEQUENCE_LENGTH,
+        truncation=True,
+        add_special_tokens=False,
+    )
+
+
+ds = ds.map(tokenize, remove_columns=ds.column_names)
+
+# define a llmcompressor recipe for FP8 W8A8 quantization
+# since the MoE gate layers are sensitive to quantization, we add them to the ignore
+# list so they remain at full precision
+recipe = [
+    QuantizationModifier(
+        targets="Linear",
+        scheme="FP8",
+        ignore=["lm_head", "re:.*mlp.gate$"],
+    ),
+]
+
+SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8"
+
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    save_compressed=True,
+    output_dir=SAVE_DIR,
+)
+
+
+print("========== SAMPLE GENERATION ==============")
+SAMPLE_INPUT = ["I love quantization because"]
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+inputs = tokenizer(SAMPLE_INPUT, return_tensors="pt", padding=True).to(model.device)
+output = model.generate(**inputs, max_length=50)
+text_output = tokenizer.batch_decode(output)
+print(text_output)

examples/quantizing_moe/deepseek_moe_w8a8.py renamed to examples/quantizing_moe/deepseek_moe_w8a8_int8.py

@@ -62,7 +62,7 @@ def tokenize(sample):
 
 ds = ds.map(tokenize, remove_columns=ds.column_names)
 
-# define a llmcompressor recipe for W416 quantization
+# define a llmcompressor recipe for INT8 W8A8 quantization
 # since the MoE gate layers are sensitive to quantization, we add them to the ignore
 # list so they remain at full precision
 recipe = [

examples/quantizing_moe/mixtral_moe_fp8.py renamed to examples/quantizing_moe/mixtral_moe_w8a8_fp8.py

@@ -2,7 +2,7 @@
 
 from transformers import AutoTokenizer
 
-from llmcompressor.modifiers.quantization.gptq import GPTQModifier
+from llmcompressor.modifiers.quantization import QuantizationModifier
 from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
 from llmcompressor.transformers.compression.helpers import calculate_offload_device_map
 
@@ -34,9 +34,7 @@
     "re:.*block_sparse_moe.gate",  # does not quantize well
 ]
 
-recipe = GPTQModifier(
-    scheme="FP8", targets="Linear", ignore=layers_to_ignore, sequential_update=True
-)
+recipe = QuantizationModifier(scheme="FP8", targets="Linear", ignore=layers_to_ignore)
 
 
 oneshot(