add rwkv support

configs/neox_arguments.md

@@ -111,7 +111,7 @@ Logging Arguments
 
 - **git_hash**: str
 
-    Default = 11a5537
+    Default = 7550d64
 
     current git hash of repository
 
@@ -432,7 +432,7 @@ Model Arguments
     The first item in the list specifies the attention type(s), and should be a list of strings. The second item
     specifies the number of times to repeat those attention types in the full list.
 
-    attention type choices:  [global, local, sparse_fixed, sparse_variable, bslongformer, bigbird, "gmlp", "amlp", "flash", "mamba"]
+    attention type choices:  [global, local, sparse_fixed, sparse_variable, bslongformer, bigbird, "gmlp", "amlp", "flash", "mamba", "rwkv"]
 
     So a 12 layer network with only global attention could be specified like:
         [[[`global`], 12]]
@@ -1965,7 +1965,9 @@ Args for deepspeed config
 
     Default = None
 
-    Configuration for using bfloat16 floating-point format as an alternative to FP16. BFLOAT16 requires hardware support (e.g., NVIDIA A100). Dictionary options as described in Deepspeed documentation: https://www.deepspeed.ai/docs/config-json/#bfloat16-training-options
+    Configuration for using bfloat16 floating-point format as an alternative to FP16. BFLOAT16 requires hardware support (e.g., NVIDIA A100).
+
+    Dictionary options as described in Deepspeed documentation: https://www.deepspeed.ai/docs/config-json/#bfloat16-training-options
 
 
 

configs/rwkv/170M.yml

@@ -0,0 +1,102 @@
+{
+  # Parallelism is not yet supported for rwkv
+  "pipe_parallel_size": 0,
+  "model_parallel_size": 1,
+
+  "num_layers": 12,
+  "hidden_size": 768,
+  "num_attention_heads": 12, # head_size = dim_att / num_attention_heads.
+                             # head_size is 64 for all rwkv models
+  "seq_length": 512,
+  "max_position_embeddings": 2048,
+  "output_layer_parallelism": "column",
+  "norm": "rmsnorm",
+  "rms_norm_epsilon": 1.0e-5,
+  "train_micro_batch_size_per_gpu": 1,
+
+  "attention_config": [[["rwkv"], 12]],
+
+  "activation": "silu",
+
+  # model settings
+
+  #"pos_emb": "rotary",
+  "rotary_pct": 0.25,
+  "no_weight_tying": true,
+  "gpt_j_residual": true,
+
+  # these should provide some speedup but takes a while to build, set to true if desired
+  "scaled_upper_triang_masked_softmax_fusion": false,
+  "bias_gelu_fusion": false,
+  "rope_fusion": false,
+  "layernorm_fusion": false,
+
+
+  # init methods
+  "init_method": "small_init",
+  "output_layer_init_method": "wang_init",
+
+  # optimizer settings
+  "optimizer": {
+    "type": "Adam",
+    "params": {
+      "lr": 0.0008,
+      "betas": [0.9, 0.95],
+      "eps": 1.0e-8,
+    }
+  },
+  "min_lr": 0.00008,
+
+  # for all zero_optimization options, see https://www.deepspeed.ai/docs/config-json/#zero-optimizations-for-fp16-training
+  "zero_optimization": {
+    "stage": 1,
+    "allgather_partitions": True,
+    "allgather_bucket_size": 500000000,
+    "overlap_comm": True,
+    "reduce_scatter": True,
+    "reduce_bucket_size": 500000000,
+    "contiguous_gradients": True,
+  },
+
+  # batch / data settings
+  "data_impl": "mmap",
+  "num_workers": 1,
+
+  # activation checkpointing
+  "checkpoint_activations": true,
+  "checkpoint_num_layers": 1,
+  "partition_activations": true,
+  "synchronize_each_layer": true,
+
+  # regularization
+  "gradient_clipping": 1.0,
+  "weight_decay": 0.1,
+  "hidden_dropout": 0,
+  "attention_dropout": 0,
+
+  # precision settings
+  "bf16": {
+    "bf16": true,
+    "enabled": true,
+    "loss_scale": 0,
+    "loss_scale_window": 1000,
+    "initial_scale_power": 12,
+    "hysteresis": 2,
+    "min_loss_scale": 1,
+  },
+
+  # misc. training settings
+  "train_iters": 500,
+  "lr_decay_iters": 500,
+  "distributed_backend": "nccl",
+  "lr_decay_style": "constant",
+  "warmup": 0.01,
+  "checkpoint_factor": 100,
+  "eval_interval": 100000,
+  "eval_iters": 10,
+
+  # logging
+  "log_interval": 10,
+  "steps_per_print": 10,
+  "wall_clock_breakdown": true,
+}

megatron/model/gpt2_model.py

@@ -37,6 +37,7 @@
     ParallelLinear,
 )
 from megatron.model.gmlp import GMLPBlock
+from megatron.model.rwkv import RWKVResidualLayerPipe
 from megatron.model.mamba import ParallelMambaResidualLayerPipe
 from megatron.model.word_embeddings import EmbeddingPipe, SoftEmbedding
 
@@ -175,6 +176,7 @@ def insert_layers(
                 "GMLPBlock",
                 "ParallelTransformerLayerPipe",
                 "ParallelMambaResidualLayerPipe",
+                "RWKVResidualLayerPipe",
             ],
         )
 
@@ -251,6 +253,14 @@ def init_specs(self):
                         mask_fn=gpt2_attention_mask_func,
                     )
                 )
+            elif layer_type == "rwkv":
+                self.specs.append(
+                    LayerSpec(
+                        RWKVResidualLayerPipe,
+                        neox_args=self.neox_args,
+                        layer_number=i,
+                    )
+                )
             elif layer_type in ["mamba"]:
                 self.specs.append(
                     LayerSpec(

megatron/model/rwkv/__init__.py

@@ -0,0 +1 @@
+from .rwkv import RWKVResidualLayerPipe, RWKVResidualLayer