Upgrade to Transformers 4.43

examples/audio-classification/run_audio_classification.py

@@ -46,8 +46,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.14.0", "To fix: pip install -r examples/pytorch/audio-classification/requirements.txt")
 

examples/contrastive-image-text/clip_media_pipe.py

@@ -31,7 +31,6 @@
         media_ext_reader_op_impl,
         media_ext_reader_op_tensor_info,
     )
-    from habana_frameworks.torch.hpu import get_device_name
 except ImportError:
     pass
 
@@ -47,7 +46,7 @@ class read_image_text_from_dataset(media_ext_reader_op_impl):
 
     """
 
-    def __init__(self, params):
+    def __init__(self, params, fw_params):
         self.batch_size = 1
         params = params["priv_params"]
         self.meta_dtype = params["label_dtype"]
@@ -64,9 +63,7 @@ def __init__(self, params):
         else:
             self.max_file = get_max_file([img["path"] for img in self.dataset["image"]])
         logger.info(f"The largest file is {self.max_file}.")
-
-    def set_params(self, params):
-        self.batch_size = params.batch_size
+        self.batch_size = fw_params.batch_size
 
     def gen_output_info(self):
         out_info = []
@@ -134,7 +131,7 @@ class ClipMediaPipe(MediaPipe):
     instance_count = 0
 
     def __init__(self, dataset=None, sampler=None, batch_size=512, drop_last=False, queue_depth=1):
-        self.device = get_device_name()
+        self.device = "legacy"
         self.dataset = dataset
         self.drop_last = drop_last
         self.sampler = sampler
@@ -157,7 +154,7 @@ def __init__(self, dataset=None, sampler=None, batch_size=512, drop_last=False,
         def_output_image_size = [self.image_size, self.image_size]
         res_pp_filter = ftype.BICUBIC
         self.decode = fn.ImageDecoder(
-            device=self.device,
+            device="hpu",
             output_format=imgtype.RGB_P,
             random_crop_type=randomCropType.CENTER_CROP,
             resize=def_output_image_size,

examples/contrastive-image-text/run_bridgetower.py

@@ -56,8 +56,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/contrastive-image-text/requirements.txt")
 

examples/contrastive-image-text/run_clip.py

@@ -61,8 +61,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/contrastive-image-text/requirements.txt")
 

examples/image-classification/README.md

@@ -45,7 +45,7 @@ python run_image_classification.py \
     --num_train_epochs 5 \
     --per_device_train_batch_size 128 \
     --per_device_eval_batch_size 64 \
-    --evaluation_strategy epoch \
+    --eval_strategy epoch \
     --save_strategy epoch \
     --load_best_model_at_end True \
     --save_total_limit 3 \
@@ -197,7 +197,7 @@ python ../gaudi_spawn.py \
     --num_train_epochs 5 \
     --per_device_train_batch_size 128 \
     --per_device_eval_batch_size 64 \
-    --evaluation_strategy epoch \
+    --eval_strategy epoch \
     --save_strategy epoch \
     --load_best_model_at_end True \
     --save_total_limit 3 \
@@ -237,7 +237,7 @@ python ../gaudi_spawn.py \
     --num_train_epochs 5 \
     --per_device_train_batch_size 128 \
     --per_device_eval_batch_size 64 \
-    --evaluation_strategy epoch \
+    --eval_strategy epoch \
     --save_strategy epoch \
     --load_best_model_at_end True \
     --save_total_limit 3 \

examples/image-classification/run_image_classification.py

@@ -63,8 +63,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/image-classification/requirements.txt")
 

examples/image-to-text/README.md

@@ -136,9 +136,9 @@ QUANT_CONFIG=./quantization_config/maxabs_quant.json python run_pipeline.py \
 
 ### Inference with FusedSDPA
 
-Habana FusedSDPA is a fused and optimized implementation of torch.nn.functional.scaled_dot_product_attention() for Gaudi. For more details, refer to [Gaudi online documentation](https://docs.habana.ai/en/latest/PyTorch/Model_Optimization_PyTorch/Optimization_in_PyTorch_Models.html?highlight=fusedsdpa#using-fused-scaled-dot-product-attention-fusedsdpa). Currently FusedSDPA works with BF16 precision for Llava models.
+Habana FusedSDPA is a fused and optimized implementation of torch.nn.functional.scaled_dot_product_attention() for Gaudi. For more details, refer to [Gaudi online documentation](https://docs.habana.ai/en/latest/PyTorch/Model_Optimization_PyTorch/Optimization_in_PyTorch_Models.html?highlight=fusedsdpa#using-fused-scaled-dot-product-attention-fusedsdpa).
 
-Use the following commands to run Llava-1.5-7b inference with FusedSDPA
+Use the following command to run Llava-1.5-7b BF16 inference with FusedSDPA
 ```bash
 python3 run_pipeline.py \
     --model_name_or_path llava-hf/llava-1.5-7b-hf \
@@ -149,12 +149,33 @@ python3 run_pipeline.py \
 ```
 
 
-Use the following commands to run Llava-v1.6-mistral-7b inference with FusedSDPA
+Use the following command to run Llava-v1.6-mistral-7b BF16 inference with FusedSDPA
 ```bash
 python3 run_pipeline.py \
     --model_name_or_path llava-hf/llava-v1.6-mistral-7b-hf \
     --image_path "https://llava-vl.github.io/static/images/view.jpg" \
     --use_hpu_graphs \
     --bf16 \
     --use_flash_attention
-```
+```
+
+
+Use the following commands to run Llava-v1.6-mistral-7b FP8 inference with FusedSDPA
+
+Here is an example of measuring the tensor quantization statistics on Llava-v1.6-mistral-7b:
+```bash
+QUANT_CONFIG=./quantization_config/maxabs_measure.json python run_pipeline.py \
+--model_name_or_path llava-hf/llava-v1.6-mistral-7b-hf \
+--image_path "https://llava-vl.github.io/static/images/view.jpg" \
+--use_hpu_graphs \
+--bf16 --use_flash_attention
+```
+
+Here is an example of quantizing the model based on previous measurements for Llava-v1.6-mistral-7b:
+```bash
+QUANT_CONFIG=./quantization_config/maxabs_quant.json python run_pipeline.py \
+--model_name_or_path llava-hf/llava-v1.6-mistral-7b-hf \
+--image_path "https://llava-vl.github.io/static/images/view.jpg" \
+--use_hpu_graphs \
+--bf16 --use_flash_attention
+```

examples/language-modeling/README.md

@@ -377,7 +377,7 @@ python3 run_lora_clm.py \
     --output_dir ./model_lora_llama \
     --num_train_epochs 3 \
     --per_device_train_batch_size 16 \
-    --evaluation_strategy "no" \
+    --eval_strategy "no" \
     --save_strategy "no" \
     --learning_rate 1e-4 \
     --warmup_ratio  0.03 \
@@ -410,7 +410,7 @@ LOWER_LIST=ops_bf16.txt python3 run_lora_clm.py \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
     --gradient_accumulation_steps 16 \
-    --evaluation_strategy "no" \
+    --eval_strategy "no" \
     --save_strategy "no" \
     --learning_rate 3e-4 \
     --max_grad_norm  0.3 \
@@ -445,7 +445,7 @@ python ../gaudi_spawn.py \
     --num_train_epochs 3 \
     --per_device_train_batch_size 8 \
     --gradient_accumulation_steps 2 \
-    --evaluation_strategy "no" \
+    --eval_strategy "no" \
     --save_strategy "no" \
     --learning_rate 3e-4 \
     --warmup_ratio  0.03 \
@@ -480,7 +480,7 @@ LOWER_LIST=ops_bf16.txt python ../gaudi_spawn.py \
 	--num_train_epochs 3 \
 	--per_device_train_batch_size 16 \
 	--gradient_accumulation_steps 1 \
-	--evaluation_strategy "no" \
+	--eval_strategy "no" \
 	--save_strategy "no" \
 	--learning_rate 3e-4 \
 	--warmup_ratio 0.03 \
@@ -518,7 +518,7 @@ python ../gaudi_spawn.py \
     --num_train_epochs 5 \
     --per_device_train_batch_size 4 \
     --per_device_eval_batch_size 4 \
-    --evaluation_strategy "no" \
+    --eval_strategy "no" \
     --save_strategy "no" \
     --learning_rate 1e-4 \
     --logging_steps 1 \
@@ -547,7 +547,7 @@ LOWER_LIST=ops_bf16.txt python3 ../gaudi_spawn.py \
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
     --gradient_accumulation_steps 16 \
-    --evaluation_strategy "no" \
+    --eval_strategy "no" \
     --save_strategy "no" \
     --learning_rate 4e-4 \
     --max_grad_norm  0.3 \
@@ -589,7 +589,7 @@ python3 ../gaudi_spawn.py --use_deepspeed  --world_size 8  run_lora_clm.py \
   --per_device_train_batch_size 10 \
   --per_device_eval_batch_size 1 \
   --gradient_checkpointing \
-  --evaluation_strategy epoch \
+  --eval_strategy epoch \
   --eval_delay 2 \
   --save_strategy no \
   --learning_rate 0.0018 \
@@ -641,7 +641,7 @@ python3 ../gaudi_spawn.py --world_size 8 --use_mpi run_lora_clm.py \
   --fsdp_config fsdp_config.json \
   --fsdp auto_wrap \
   --num_train_epochs 2 \
-  --evaluation_strategy epoch \
+  --eval_strategy epoch \
   --per_device_eval_batch_size 1 \
   --eval_delay 2 \
   --do_eval \
@@ -668,7 +668,7 @@ DEEPSPEED_HPU_ZERO3_SYNC_MARK_STEP_REQUIRED=1 LOWER_LIST=ops_bf16.txt python3 ..
     --per_device_train_batch_size 1 \
     --per_device_eval_batch_size 1 \
     --gradient_accumulation_steps 16 \
-    --evaluation_strategy "no" \
+    --eval_strategy "no" \
     --save_strategy "no" \
     --learning_rate 4e-4 \
     --max_grad_norm  0.3 \

examples/language-modeling/requirements.txt

@@ -4,4 +4,4 @@ sentencepiece != 0.1.92
 protobuf
 evaluate
 scikit-learn
-peft == 0.10.0
+peft == 0.12.0

examples/language-modeling/run_clm.py

@@ -62,8 +62,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 

examples/language-modeling/run_mlm.py

@@ -61,8 +61,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 

examples/protein-folding/README.md

@@ -66,7 +66,7 @@ python ../gaudi_spawn.py --world_size 8 --use_mpi run_sequence_classification.py
     --num_train_epochs 100 \
     --lr_scheduler_type constant \
     --do_eval \
-    --evaluation_strategy epoch \
+    --eval_strategy epoch \
     --per_device_eval_batch_size 32 \
     --logging_strategy epoch \
     --save_strategy epoch \

examples/protein-folding/run_esmfold.py

@@ -40,7 +40,7 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Optimum Habana is not installed. Remove at your own risks.
-check_optimum_habana_min_version("1.11.0")
+check_optimum_habana_min_version("1.12.0")
 
 
 def convert_outputs_to_pdb(outputs):

examples/protein-folding/run_sequence_classification.py

@@ -41,7 +41,7 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Optimum Habana is not installed. Remove at your own risks.
-check_optimum_habana_min_version("1.11.0")
+check_optimum_habana_min_version("1.12.0")
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)

examples/protein-folding/run_zero_shot_eval.py

@@ -36,7 +36,7 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Optimum Habana is not installed. Remove at your own risks.
-check_optimum_habana_min_version("1.11.0")
+check_optimum_habana_min_version("1.12.0")
 
 
 logging.basicConfig(

examples/question-answering/run_qa.py

@@ -60,8 +60,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/question-answering/run_seq2seq_qa.py

@@ -56,8 +56,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/speech-recognition/README.md

@@ -237,7 +237,7 @@ python run_speech_recognition_seq2seq.py \
     --logging_steps="25" \
     --learning_rate="1e-5" \
     --warmup_steps="500" \
-    --evaluation_strategy="steps" \
+    --eval_strategy="steps" \
     --eval_steps="1000" \
     --save_strategy="steps" \
     --save_steps="1000" \

examples/speech-recognition/run_speech_recognition_ctc.py

@@ -59,8 +59,8 @@ def check_optimum_habana_min_version(*a, **b):
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
 

examples/speech-recognition/run_speech_recognition_seq2seq.py

@@ -55,8 +55,8 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.40.0")
-check_optimum_habana_min_version("1.11.0")
+check_min_version("4.43.0")
+check_optimum_habana_min_version("1.12.0")
 
 require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
 

examples/stable-diffusion/image_to_image_generation.py

@@ -40,7 +40,7 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Optimum Habana is not installed. Remove at your own risks.
-check_optimum_habana_min_version("1.10.0")
+check_optimum_habana_min_version("1.12.0")
 
 
 logger = logging.getLogger(__name__)

examples/stable-diffusion/image_to_video_generation.py

@@ -34,7 +34,7 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Optimum Habana is not installed. Remove at your own risks.
-check_optimum_habana_min_version("1.8.1")
+check_optimum_habana_min_version("1.12.0")
 
 
 logger = logging.getLogger(__name__)

examples/stable-diffusion/text_to_image_generation.py

@@ -39,7 +39,7 @@ def check_optimum_habana_min_version(*a, **b):
 
 
 # Will error if the minimal version of Optimum Habana is not installed. Remove at your own risks.
-check_optimum_habana_min_version("1.11.0")
+check_optimum_habana_min_version("1.12.0")
 
 
 logger = logging.getLogger(__name__)