Pin bump jan 26

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-7a79b41e29a790ebb4b530eb98a89381e2d7de29
+fb5eb9107d1e94001d80b4beda785aa6864d12ab

backends/openvino/requirements.txt

@@ -1,2 +1,2 @@
 transformers
-git+https://github.com/openvinotoolkit/nncf@3d753ac#egg=nncf
+git+https://github.com/openvinotoolkit/nncf@61ea196235f8edf474d19acf237f2511f3c75348

examples/models/llama/install_requirements.sh

@@ -11,7 +11,8 @@
 # Install snakeviz for cProfile flamegraph
 # Install lm-eval for Model Evaluation with lm-evalution-harness.
 # Install safetensors to load safetensors checkpoints (currently adapter only).
-pip install hydra-core huggingface_hub tiktoken torchtune sentencepiece tokenizers snakeviz lm_eval==0.4.5 blobfile safetensors
+pip install hydra-core huggingface_hub tiktoken sentencepiece tokenizers snakeviz lm_eval==0.4.5 blobfile safetensors
+pip install git+https://github.com/pytorch/torchtune.git@44271b570af36cfda8ee20a4479d2652770378c0
 
 # Call the install helper for further setup
 python examples/models/llama/install_requirement_helper.py

examples/qualcomm/oss_scripts/llama/eval_llama_qnn.py

@@ -10,7 +10,6 @@
 
 import logging
 import sys
-import types
 
 import torch
 
@@ -57,8 +56,6 @@
     QuantizationRecipe,
     quantize_module_swap,
 )
-from torchao.prototype.spinquant import apply_spinquant
-
 from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
 from torchao.quantization.pt2e.quantizer import QuantizationSpec
 from transformers import AutoTokenizer
@@ -220,24 +217,6 @@ def prequant_algorithm(model, prefill_config, args):
 
     scales_state_dict = {}
 
-    if args.spinquant:
-        config = types.SimpleNamespace(
-            dim=prefill_config.dim,
-            head_dim=prefill_config.dim // prefill_config.n_heads,
-            n_local_heads=prefill_config.n_heads,
-            intermediate_size=4 * prefill_config.dim,
-        )
-        model.config = config
-        apply_spinquant(
-            model,
-            use_r1=True,
-            use_r2=True,
-            use_r4=False,
-            pretrained_rotation_path=None,
-            qkv_split=True,
-        )
-        logging.info("Applied SpinQuant to the model")
-
     if args.range_setting == "mse_with_act_loss":
         wrapped_model = WrappedLlamaModel(
             model, *atten_mask, args.use_kv_cache, args.max_seq_length, args.device
@@ -398,11 +377,6 @@ def main() -> None:
         help="Choose which range setting method for weight quantization (e.g. mse_weight_only or mse_with_act_loss). If not specified, defaults to minmax",
         type=str,
     )
-    parser.add_argument(
-        "--spinquant",
-        help="Apply SpinQuant (R1+R2) to the model. Uses random Hadamard matrices for rotations",
-        action="store_true",
-    )
     parser.add_argument(
         "--fraction",
         help="the fraction of examples per task (only use this for testing)",

install_requirements.py

@@ -133,7 +133,7 @@ def install_optional_example_requirements(use_pytorch_nightly):
             else "torchvision"
         ),
         (
-            f"torchaudio==2.10.0.{NIGHTLY_VERSION}"
+            f"torchaudio==2.11.0.{NIGHTLY_VERSION}"
             if use_pytorch_nightly
             else "torchaudio"
         ),

requirements-examples.txt

@@ -3,5 +3,5 @@
 datasets == 3.6.0 # 4.0.0 deprecates trust_remote_code and load scripts. For now pin to 3.6.0
 timm == 1.0.7
 torchsr == 1.0.4
-torchtune >= 0.6.1
+torchtune @ git+https://github.com/pytorch/torchtune.git@44271b570af36cfda8ee20a4479d2652770378c0
 transformers == 5.0.0rc1

torch_pin.py

@@ -1,2 +1,2 @@
 TORCH_VERSION = "2.11.0"
-NIGHTLY_VERSION = "dev20251222"
+NIGHTLY_VERSION = "dev20260120"