Updated storage and training files to enable text generation tasks. A…

benchmark/mnist/mnist.yaml

@@ -12,6 +12,7 @@ model_storage:
 training:
   gpus: 1
   device: "cuda:0"
+  generative: False
   dataloader_workers: 2
   use_previous_model: True
   initial_model: random

benchmark/wildtime_benchmarks/example_pipelines/arxiv.yaml

@@ -17,6 +17,7 @@ training:
   initial_model: random
   batch_size: 128
   shuffle: True
+  generative: False
   optimizers:
     - name: "default"
       algorithm: "SGD"

benchmark/wildtime_benchmarks/example_pipelines/data_drift_trigger/arxiv_datadrift.yaml

@@ -13,6 +13,7 @@ training:
   gpus: 1
   device: "cuda:0"
   dataloader_workers: 2
+  generative: False
   use_previous_model: True
   initial_model: random
   batch_size: 96

benchmark/wildtime_benchmarks/example_pipelines/data_drift_trigger/huffpost_datadrift.yaml

@@ -13,6 +13,7 @@ training:
   gpus: 1
   device: "cuda:0"
   dataloader_workers: 2
+  generative: False
   use_previous_model: True
   initial_model: random
   batch_size: 64

benchmark/wildtime_benchmarks/example_pipelines/data_drift_trigger/yearbook_datadrift.yaml

@@ -14,6 +14,7 @@ training:
   gpus: 1
   device: "cuda:0"
   dataloader_workers: 2
+  generative: False
   use_previous_model: True
   initial_model: random
   batch_size: 64

benchmark/wildtime_benchmarks/example_pipelines/fmow.yaml

@@ -13,6 +13,8 @@ training:
   gpus: 1
   device: "cuda:0"
   dataloader_workers: 2
+  generative: False
+
   use_previous_model: True
   initial_model: random
   batch_size: 64

benchmark/wildtime_benchmarks/example_pipelines/huffpost.yaml

@@ -12,6 +12,7 @@ model_storage:
 training:
   gpus: 1
   device: "cuda:0"
+  generative: False
   dataloader_workers: 2
   use_previous_model: True
   initial_model: random

benchmark/wildtime_benchmarks/example_pipelines/yearbook.yaml

@@ -14,6 +14,7 @@ training:
   gpus: 1
   device: "cuda:0"
   dataloader_workers: 2
+  generative: False
   use_previous_model: True
   initial_model: random
   batch_size: 64

environment.yml

@@ -30,7 +30,7 @@ dependencies:
   - psycopg2
   - sqlalchemy>=2.0
   - pyaml
-  - pydantic
+  - pydantic==2.9.2
   - numpy==1.26.*
   - pandas
   - bitstring
@@ -43,11 +43,10 @@ dependencies:
   - nltk
   - pytorch::pytorch=2.2.1
   - pytorch::torchvision
-  - pytorch::cpuonly # comment out if commenting in lines below for CUDA
-#  - pytorch::pytorch-cuda=12.1
-#  - nvidia::cuda-libraries-dev=12.1.*
-#  - nvidia::cuda-nvcc=12.1.*
-#  - nvidia::cuda-nvtx=12.1.*
-#  - nvidia::cuda-cupti=12.1.*
-#  - nvidia::cuda-cudart-dev=12.1.*
-#  - nvidia::cuda-profiler-api=12.1.*
+  - pytorch::pytorch-cuda=12.1
+  - nvidia::cuda-libraries-dev=12.1.*
+  - nvidia::cuda-nvcc=12.1.*
+  - nvidia::cuda-nvtx=12.1.*
+  - nvidia::cuda-cupti=12.1.*
+  - nvidia::cuda-cudart-dev=12.1.*
+  - nvidia::cuda-profiler-api==12.1.*

integrationtests/config/dummy.yaml

@@ -12,6 +12,7 @@ model_storage:
 training:
   gpus: 1
   device: "cpu"
+  generative: False
   dataloader_workers: 1
   use_previous_model: True
   initial_model: random

integrationtests/config/rho_loss.yaml

@@ -13,6 +13,7 @@ training:
   gpus: 1
   device: "cpu"
   dataloader_workers: 2
+  generative: False
   use_previous_model: False
   initial_model: random
   batch_size: 4
@@ -60,6 +61,7 @@ selection_strategy:
       il_model_config:
         num_classes: 10
       device: "cpu"
+      generative: False
       dataloader_workers: 1
       use_previous_model: False
       batch_size: 2
@@ -75,4 +77,4 @@ selection_strategy:
                 lr: 0.1
                 momentum: 0.001
       optimization_criterion:
-        name: "CrossEntropyLoss"
+        name: "CrossEntropyLoss"

modyn/common/grpc/grpc_helpers.py

@@ -251,6 +251,7 @@ def prepare_start_training_request(
             enable_accurate_gpu_measurements=training_config.enable_accurate_gpu_measurements,
             record_loss_every=training_config.record_loss_every,
             drop_last_batch=training_config.drop_last_batch,
+            generative=training_config.generative,
         )
 
     def start_training(

modyn/config/examples/modyn_config.yaml

@@ -278,7 +278,7 @@ selector:
   local_storage_directory: "/tmp/local_storage"
   local_storage_max_samples_in_file: 1000000
   cleanup_storage_directories_after_shutdown: true
-  ignore_existing_trigger_samples: false
+  ignore_existing_trigger_samples: true
 
 trainer_server:
   hostname: "trainer_server"

modyn/config/schema/pipeline/training/config.py

@@ -119,6 +119,11 @@ class TrainingConfig(ModynBaseModel):
             "we start with random weights. If initial_model is 'pretrained', cannot be False."
         )
     )
+    generative: bool = Field(False,
+        description=(
+            "If True then, then the training pipeline goes into the generative branch, data is sampled without expecting labels."
+        )
+    )
     seed: int | None = Field(
         None,
         description=(

modyn/config/schema/system/config.py

@@ -255,7 +255,7 @@ class SelectorConfig(HostnamePortMixin):
         ),
     )
     ignore_existing_trigger_samples: bool = Field(
-        False,
+        True,
         description=(
             "Whether to ignore existing trigger samples when starting the selector. If set to false, the trigger "
             "sample directory has to be empty upon startup. May lead to unexpected behaviour if set to true and the "