Update ipu_config.json

ipu_config.json

@@ -1,20 +1,21 @@
 {
     "decompose_grad_sum": false,
-    "device_iterations": 16,
-    "embedding_serialization_factor": 1,
+    "device_iterations": 1,
+    "embedding_serialization_factor": 0,
     "enable_half_first_order_momentum": true,
-    "enable_half_partials": false,
+    "enable_half_partials": true,
     "executable_cache_dir": "./exe_cache",
     "execute_encoder_on_cpu_for_generation": false,
-    "gradient_accumulation_steps": 8,
-    "inference_device_iterations": 4,
-    "inference_replication_factor": {"pod4": 4, "pod8": 8, "pod16": 16, "pod32": 32, "pod64": 64, "default": 4},
-    "ipus_per_replica": 1,
-    "layers_per_ipu": [6],
+    "gradient_accumulation_steps": 16,
+    "inference_device_iterations": 5,
+    "inference_replication_factor": {"pod4": 2, "pod8": 4, "pod16": 8, "pod32": 16, "pod64": 32, "default": 2},
+    "ipus_per_replica": 2,
+    "layers_per_ipu": [2, 4],
+    "matmul_proportion": 0.25,
     "optimizer_state_offchip": false,
     "profile_dir": "",
-    "recompute_checkpoint_every_layer": false,
-    "replication_factor": {"pod4": 4, "pod8": 8, "pod16": 16, "pod32": 32, "pod64": 64, "default": 4},
+    "recompute_checkpoint_every_layer": true,
+    "replication_factor": {"pod4": 2, "pod8": 4, "pod16": 8, "pod32": 16, "pod64": 32, "default": 2},
     "replicated_tensor_sharding": true,
     "seed": 42,
     "sharded_execution_for_inference": false,