Merge branch 'main' into migrate-existing-logger-framework

.github/workflows/test-check.yaml

@@ -62,7 +62,7 @@ jobs:
     steps:
       - uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: '3.11'
       - uses: actions/checkout@v2
       - uses: actions/checkout@v2
         with:
@@ -86,7 +86,7 @@ jobs:
     steps:
       - uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: '3.11'
       - uses: actions/checkout@v2
       - uses: actions/checkout@v2
         with:
@@ -110,7 +110,7 @@ jobs:
     steps:
       - uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: '3.11'
       - uses: actions/checkout@v2
       - uses: actions/checkout@v2
         with:
@@ -134,7 +134,7 @@ jobs:
     steps:
       - uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: '3.11'
       - uses: actions/checkout@v2
       - uses: actions/checkout@v2
         with:

DEVELOPING.md

@@ -16,7 +16,7 @@ limitations under the License.
 
 # Developing SparseML
 
-SparseML is developed and tested using Python 3.8-3.10.
+SparseML is developed and tested using Python 3.8-3.11.
 To develop SparseML, you will also need the development dependencies and to follow the styling guidelines.
 
 Here are some details to get started.

README.md

@@ -111,7 +111,7 @@ SparseML enables you to create a sparse model trained on your dataset in two way
 
 ## Installation
 
-This repository is tested on Python 3.8-3.10, and Linux/Debian systems.
+This repository is tested on Python 3.8-3.11, and Linux/Debian systems.
 
 It is recommended to install in a [virtual environment](https://docs.python.org/3/library/venv.html) to keep your system in order.
 Currently supported ML Frameworks are the following: `torch>=1.1.0,<=2.0`, `tensorflow>=1.8.0,<2.0.0`, `tensorflow.keras >= 2.2.0`.

docs/source/installation.md

@@ -16,7 +16,7 @@ limitations under the License.
 
 # Installation
 
-This repository is tested on Python 3.8-3.10, and Linux/Debian systems.
+This repository is tested on Python 3.8-3.11, and Linux/Debian systems.
 It is recommended to install in a [virtual environment](https://docs.python.org/3/library/venv.html) to keep your system in order.
 Currently supported ML Frameworks are the following: `torch>=1.1.0,<1.14`, `tensorflow>=1.8.0,<=2.0.0`, `tensorflow.keras >= 2.2.0`.
 

setup.py

@@ -293,7 +293,7 @@ def _setup_long_description() -> Tuple[str, str]:
     install_requires=_setup_install_requires(),
     extras_require=_setup_extras(),
     entry_points=_setup_entry_points(),
-    python_requires=">=3.8.0,<3.11",
+    python_requires=">=3.8.0,<3.12",
     classifiers=[
         "Development Status :: 5 - Production/Stable",
         "Programming Language :: Python :: 3",

src/sparseml/core/lifecycle/session.py

@@ -31,7 +31,7 @@
 @dataclass
 class SparsificationLifecycle:
     state: Optional[State] = None
-    recipe_container: RecipeContainer = RecipeContainer()
+    recipe_container: RecipeContainer = field(default_factory=RecipeContainer)
     modifiers: List[ModifierInterface] = field(default_factory=list)
     event_lifecycle: Optional[EventLifecycle] = None
 

src/sparseml/experimental/sparsegpt/dispatch.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-SUPPORTED_MODELS = ["opt", "mpt", "llama"]
+SUPPORTED_MODELS = ["opt", "mpt", "llama-2"]
 
 
 def load_model(args, model_key: str = None, *gargs, **kwargs):
@@ -21,7 +21,7 @@ def load_model(args, model_key: str = None, *gargs, **kwargs):
         from sparseml.experimental.sparsegpt.opt import load_model as _load_model
     elif model_key == "mpt":
         from sparseml.experimental.sparsegpt.mpt import load_model as _load_model
-    elif model_key == "llama":
+    elif model_key == "llama-2":
         from sparseml.experimental.sparsegpt.llama2 import load_model as _load_model
     else:
         raise ValueError(f"Unrecognized model key. Supported: {SUPPORTED_MODELS}")
@@ -34,7 +34,7 @@ def load_data(args, model_key: str = None, *gargs, **kwargs):
         from sparseml.experimental.sparsegpt.opt import load_data as _load_data
     elif model_key == "mpt":
         from sparseml.experimental.sparsegpt.mpt import load_data as _load_data
-    elif model_key == "llama":
+    elif model_key == "llama-2":
         from sparseml.experimental.sparsegpt.llama2 import load_data as _load_data
     else:
         raise ValueError(f"Unrecognized model key. Supported: {SUPPORTED_MODELS}")
@@ -47,7 +47,7 @@ def evaluate_perplexity(
     model_key = _get_model_key(args) if model_key is None else model_key
     if model_key == "opt":
         from sparseml.experimental.sparsegpt.opt import ppl_eval as _ppl_eval
-    elif model_key == "llama":
+    elif model_key == "llama-2":
         from sparseml.experimental.sparsegpt.llama2 import ppl_eval as _ppl_eval
     else:
         raise ValueError(f"Unrecognized model key. Supported: {SUPPORTED_MODELS}")
@@ -64,7 +64,7 @@ def prepare_sparsegpt(model, dataloader, args, model_key: str = None, **kwargs):
         from sparseml.experimental.sparsegpt.mpt import (
             prepare_sparsegpt as _prepare_sparsegpt,
         )
-    elif model_key == "llama":
+    elif model_key == "llama-2":
         from sparseml.experimental.sparsegpt.llama2 import (
             prepare_sparsegpt as _prepare_sparsegpt,
         )

src/sparseml/experimental/sparsegpt/examples/llama2/compare_obcq.py

@@ -23,7 +23,7 @@
 
 dataset = "open_platypus"
 model_name = "/home/sadkins/ml-experiments/nlg-text_generation/"
-model_name += "llama_chat-llama_7b_chat-base/dense/training"
+model_name += "llama_chat-llama_7b_chat-base/dense_llama-2/training"
 sparsity = 0.5
 nbits = 8
 smooth_quant = 0
@@ -73,7 +73,7 @@ class ProdArgs:
 
 
 def run_experimental_obcq(experimental_args):
-    model = load_model(experimental_args)
+    model, _ = load_model(experimental_args)
     calibration_data, _, _ = load_data(experimental_args, data_sequence_length)
     sequential(model, calibration_data, device, experimental_args)
 

src/sparseml/experimental/sparsegpt/examples/opt/compare_obcq.py

@@ -72,7 +72,7 @@ class ProdArgs:
 
 
 def run_experimental_obcq(experimental_args):
-    model = load_model(experimental_args)
+    model, _ = load_model(experimental_args)
     calibration_data, _, _ = load_data(experimental_args, data_sequence_length)
     sequential(model, calibration_data, device, experimental_args)
 

src/sparseml/experimental/sparsegpt/examples/opt/scripts/prune_quantize.opt.0.sh

@@ -2,11 +2,11 @@
 
 export CUDA_VISIBLE_DEVICES=0
 
-ROOT=$HOME/sparseml/src/sparseml/experimental/sparsegpt
+ROOT=$HOME/src/neuralmagic/sparseml/src/sparseml/experimental/sparsegpt
 
 DATASET=c4
 
-RECIPE_DIR=$ROOT/examples/opt/recipes
+RECIPE_DIR=$ROOT/recipes
 RECIPE_NAME=opt-1.3b-opt_pretrain-pruned50_quantW8A8
 
 SRC_MODEL_ORG=facebook

src/sparseml/experimental/sparsegpt/layer_compressor.py

@@ -22,9 +22,6 @@
 from sparseml.experimental.sparsegpt.sparsegpt import SparseGPT
 
 
-DEFAULT_WBITS = 16
-
-
 class BaseCompressor:
     def __init__(self, model):
         self.model = model