new large gpu launch

experiments/plan/config/gpu_large_v01base.yaml

@@ -0,0 +1,82 @@
+---
+tags: []
+
+system:
+  threads: 128
+  runtime: 600
+  memory: auto
+
+data:
+  name: pypsa
+  case: load_medium
+  add_ground: true
+  use_batteries: true
+  num_nodes: 1000
+  use_extra_components: false
+  start_hour: peak_hybrid_day
+  num_hours: 8640
+  dont_expand: [hydro, coal, biomass]
+  scale_by_hours: false
+  args:
+    power_unit: 1000.0  # 1e3 = GW
+    cost_unit: 100.0  # $ / MWh
+    marginal_load_value: 500.0  # $ / MWh
+    scale_load: 1.0
+    scale_generator_capacity_factor: 0.7
+    scale_line_capacity_factor: 0.7
+    carbon_tax: 0.0  # $ / ton CO2
+    generator_cost_perturbation: 1.0  # MW
+    load_cost_perturbation: 10.0  # MW
+    drop_empty_generators: false
+    expand_empty_generators: 0.5  # MW
+    battery_discharge_cost: 1.0  # $ / MWh
+    battery_init_soc: 0.0
+    battery_final_soc: 0.0
+
+problem:
+  stochastic: true
+  hours_per_scenario: 24
+  cost_weight: 1.0
+  emissions_weight: 200.0  # $ / ton CO2
+  regularize: 1.0e-6  # Jacobian regularization
+  parameters: [generator, dc_line, ac_line, battery]
+
+optimizer:
+  name: gradient_descent
+  batch_size: 8
+  parallel: true
+  num_parallel_workers: 8
+  num_iterations: 500
+  initial_state: initial
+  use_wandb: true
+  log_wandb_every: 1
+  checkpoint_every: 1
+  track_full_loss_every: -1
+  args:
+    step_size: [expand, 1.0e-2, 1.0e-3]
+    clip: 1000.0
+
+layer:
+  use_admm: false
+  torch_dtype: float32
+  args:
+    # Settings
+    machine: cuda
+    resid_norm: 2
+    battery_window: 24
+    verbose: false
+
+    # Duration / convergence
+    num_iterations: 1000
+    minimum_iterations: 250
+    atol: 1.0e-3
+
+    # Penalty parameter
+    adaptive_rho: true
+    rho_power: 1.0
+    rho_angle: 1.0
+    alpha: 1.0
+    adaptation_frequency: 50
+
+relaxation:
+  should_solve: false

experiments/plan/config/gpu_large_v01exp.yaml

@@ -0,0 +1,81 @@
+---
+tags: []
+
+system:
+  threads: 16
+  runtime: 120
+  memory: auto
+  gpu: 1
+
+data:
+  name: pypsa
+  case: load_medium
+  add_ground: true
+  use_batteries: true
+  num_nodes: 1000
+  use_extra_components: false
+  start_hour: peak_hybrid_day
+  num_hours: 8640
+  dont_expand: [hydro, coal, biomass]
+  scale_by_hours: false
+  args:
+    power_unit: 1000.0  # 1e3 = GW
+    cost_unit: 100.0  # $ / MWh
+    marginal_load_value: 500.0  # $ / MWh
+    scale_load: 1.0
+    scale_generator_capacity_factor: 0.7
+    scale_line_capacity_factor: 0.7
+    carbon_tax: 0.0  # $ / ton CO2
+    generator_cost_perturbation: 1.0  # MW
+    load_cost_perturbation: 10.0  # MW
+    drop_empty_generators: false
+    expand_empty_generators: 0.5  # MW
+    battery_discharge_cost: 1.0  # $ / MWh
+    battery_init_soc: 0.0
+    battery_final_soc: 0.0
+
+problem:
+  stochastic: true
+  hours_per_scenario: 192
+  cost_weight: 1.0
+  emissions_weight: 200.0  # $ / ton CO2
+  regularize: 1.0e-6  # Jacobian regularization
+  parameters: [generator, dc_line, ac_line, battery]
+
+optimizer:
+  name: gradient_descent
+  batch_size: 1
+  num_iterations: 500
+  initial_state: initial
+  use_wandb: true
+  log_wandb_every: 1
+  checkpoint_every: 1
+  track_full_loss_every: -1
+  args:
+    step_size: [expand, 1.0e-2, 1.0e-3]
+    clip: 1000.0
+
+layer:
+  use_admm: true
+  torch_dtype: float32
+  args:
+    # Settings
+    machine: cuda
+    resid_norm: 2
+    battery_window: 24
+    verbose: false
+
+    # Duration / convergence
+    num_iterations: 1000
+    minimum_iterations: 250
+    atol: 1.0e-3
+
+    # Penalty parameter
+    adaptive_rho: true
+    rho_power: 1.0
+    rho_angle: 1.0
+    alpha: 1.0
+    adaptation_frequency: 50
+
+relaxation:
+  should_solve: false

experiments/plan/runner.py

@@ -774,7 +774,13 @@ def all_op_costs(J, grad, params, last_state, problem):
         if track_full_loss_every == 0:  # Track once per batch
             track_full_loss_every = int(num_problems / batch_size)
 
-        if batch_size != num_problems:
+        if (batch_size == num_problems) or (batch_size == -1):
+            print("Not tracking full loss because batch_size is 0, num_problems, or -1.")
+
+            def full_loss_tracker(J, grad, params, last_state, _stoch_prob):
+                return J
+
+        else:
             print(f"Tracking full loss every {track_full_loss_every} batches.")
 
             def full_loss_tracker(J, grad, params, last_state, _stoch_prob):
@@ -787,11 +793,6 @@ def full_loss_tracker(J, grad, params, last_state, _stoch_prob):
                     return problem.full_loss
                 else:
                     return getattr(problem, "full_loss", np.inf)
-        else:
-            print("Not tracking full loss because batch_size is 0.")
-
-            def full_loss_tracker(J, grad, params, last_state, _stoch_prob):
-                return J
     else:
 
         def full_loss_tracker(J, grad, params, last_state, problem):

experiments/solve/config/scaling_devices_v04cpu.yaml

@@ -0,0 +1,83 @@
+---
+
+system:
+  threads: 128
+  runtime: 240  # minutes
+  memory: auto
+
+data:
+  name: pypsa
+  case: load_medium
+  start_hour: peak_hybrid_day
+  num_nodes: [expand, 100, 500, 1000, 2000, 4000]
+  num_hours: 240  # Total number of hourds in the dataset
+  add_ground: true
+  use_batteries: true
+  use_extra_components: false
+  dont_expand: [hydro, coal, biomass]
+  scale_by_hours: false
+  args:
+    power_unit: 1000.0  # 1e3 = GW
+    cost_unit: 100.0  # $ / MWh
+    marginal_load_value: 500.0  # $ / MWh
+    scale_load: [expand, 0.5, 0.6, 0.7]
+    scale_generator_capacity_factor: 0.7
+    scale_line_capacity_factor: 0.7
+    carbon_tax: 0.0  # $ / ton CO2
+    generator_cost_perturbation: 1.0  # MW
+    load_cost_perturbation: 10.0  # MW
+    drop_empty_generators: false
+    expand_empty_generators: 0.5  # MW
+    battery_discharge_cost: 1.0  # $ / MWh
+    battery_init_soc: 0.0
+    battery_final_soc: 0.0
+
+parameters:
+  parameter_types: [generator, dc_line, ac_line, battery]
+  hours_per_scenario: 24  # Hours in one "batch"
+  capacities: [base]  # base, 1.5, base_v07/000/model_00500
+
+solver: admm
+
+battery_window: 24  # Hours for battery time horizon
+
+cvxpy_args:
+  solver: MOSEK
+  solver_kwargs:
+    verbose: true
+    accept_unknown: true
+    mosek_params:
+      MSK_IPAR_NUM_THREADS: 16
+      # MSK_DPAR_INTPNT_TOL_REL_GAP: 1.0e-3
+      # MSK_DPAR_INTPNT_TOL_DFEAS: 1.0e-3
+      # MSK_DPAR_INTPNT_TOL_PFEAS: 1.0e-3
+
+admm_args:
+  # Machine settings
+  machine: cpu
+  dtype: float32
+  resid_norm: 2
+  safe_mode: false
+  scale_dual_residuals: true
+  relative_rho_angle: false
+  verbose: false
+
+  # Duration and tolerance
+  num_iterations: 5000
+  minimum_iterations: 5000
+  atol: 1.0e-3
+
+  # Step size
+  rho_power: 1.0
+  rho_angle: 1.0
+  alpha: [sweep, 1.0, 1.5]  # Over-relaxation
+
+  # Adaptation
+  adaptive_rho: true
+  adaptation_frequency: 10
+  adaptation_tolerance: 2.0
+  tau: 1.1
+
+  # Battery prox operator parameters
+  battery_inner_iterations: 10
+  battery_inner_weight: 1.0

zap/layer.py

@@ -70,7 +70,7 @@ def forward(self, **kwargs) -> DispatchOutcome:
             contingency_device=self.contingency_device,
             contingency_mask=self.contingency_mask,
         )
-        print("Dispatch: ", time.time() - start)
+        # print("Dispatch: ", time.time() - start)
         return result
 
     def backward(self, z: DispatchOutcome, dz: DispatchOutcome, regularize=1e-8, **kwargs):