fast and good grad checkpoint, gelu torch script

act_ck.py

+from main import *
+import time
+import gc
+
+from time import perf_counter, perf_counter_ns
+import numpy as np
+from tqdm import tqdm
+from contextlib import contextmanager
+#replicating timeit magic function of ipython
+def timeit(func, r=1, n=5, quiet=False, function=None, do_tqdm=False, first=True):
+    precision = 'ns'
+    r_arr = np.empty([2, r]) # [0] = mean, [1] = std
+    if function:
+        func.__name__ = function.__name__
+
+    for i in tqdm(range(r)) if do_tqdm else range(r):
+        n_arr = np.empty(n)
+        for k in range(n):
+            start = perf_counter_ns()
+            func()
+            n_arr[k] = perf_counter_ns() - start
+        
+        if not first:
+            # delete the first element from n_arr numpy array
+            n_arr = np.delete(n_arr, 0)
+
+        r_arr[0, i] = np.mean(n_arr)
+        r_arr[1, i] = np.std(n_arr)
+    
+    best = r_arr[:, np.argmin(r_arr[0])] # [0] = mean, [1] = std
+    #check if best[0] bigger than 1ms in numpy
+    if best[0] < 1e3:
+        precision = 'ns'
+
+    elif best[0] >= 1e9:
+        best[0] = best[0] * 1e-9
+        best[1] = best[1] * 1e-9
+        precision = 's'
+
+    elif best[0] >= 1e6:
+        best[0] = best[0] * 1e-6
+        best[1] = best[1] * 1e-6
+        precision = 'ms'
+
+    elif best[0] >= 1e3:
+        precision = 'μs'
+        best[0] = best[0] * 1e-3
+        best[1] = best[1] * 1e-3
+
+    if not quiet:
+        if precision == 'ns':
+            print(f"{func.__name__}: {best[0]:.0f}{precision} ± {best[1]:.0f}{precision} per loop (mean ± std. dev. of {str(r)} runs, {str(n)} loops each)")
+        if precision == 'μs':
+            print(f"{func.__name__}: {best[0]:.2f}{precision} ± {best[1]:.2f}{precision} per loop (mean ± std. dev. of {str(r)} runs, {str(n)} loops each)")
+        elif precision == 'ms':
+            print(f"{func.__name__}: {best[0]:.2f}{precision} ± {best[1]:.2f}{precision} per loop (mean ± std. dev. of {str(r)} runs, {str(n)} loops each)")
+        elif precision == 's':
+            print(f"{func.__name__}: {best[0]:.4f}{precision} ± {best[1]:.4f}{precision} per loop (mean ± std. dev. of {str(r)} runs, {str(n)} loops each)")
+
+
+def rndinput(shape):
+    return torch.randint(0, 50256, shape).long().cuda()
+
+def forward(model, x):
+    out = model.get_logits(x, act_ck=False)
+    print(out.shape)
+    print("torch.cuda.memory_allocated: %fGB"%(torch.cuda.memory_allocated(0)/1024/1024/1024))
+    loss = torch.nn.CrossEntropyLoss()(out, out)
+    loss.backward()
+    model.zero_grad()
+    print("torch.cuda.memory_allocated: %fGB"%(torch.cuda.memory_allocated(0)/1024/1024/1024))
+
+def main():
+    model = init_1_3b().cuda().half()
+    shape = (1, 2048)
+    #print(model(x).shape)
+    print("PyTorch Eager")
+    timeit(r=1, n=2, func=lambda: forward(model, rndinput(shape)), do_tqdm=False, first=False)
+
+if __name__ == "__main__":
+    main()

gptj.py

@@ -60,15 +60,19 @@ def timeit(func, r=1, n=5, quiet=False, function=None, do_tqdm=False, first=True
 def rndinput(shape):
     return torch.randint(0, 50256, shape).long().cuda()
 
-with torch.no_grad():
-    model = init_6b().cuda().half()
-    shape = (1, 1)
+@torch.no_grad()
+def main():
+    model = init_1_3b().cuda().half()
+    shape = (1, 2048)
     x = torch.zeros(shape).cuda().long()
     print(model(x).shape)
     print("PyTorch Eager")
-    timeit(r=1, n=1, func=lambda: model(x), do_tqdm=False, first=True)
+    timeit(r=1, n=10, func=lambda: model(x), do_tqdm=False, first=False)
     with torch.jit.fuser("fuser2"):
         module = torch.jit.trace(model, torch.zeros(shape).long().cuda())
         torch.jit.optimize_for_inference(module)
     print("PyTorch JIT")
-    timeit(r=1, n=1, func=lambda: module(rndinput((1, 1))), do_tqdm=False, first=True)
+    timeit(r=1, n=10, func=lambda: module(rndinput(shape)), do_tqdm=False, first=False)
+
+if __name__ == "__main__":
+    main()

lm_train/models.py

+from transformers import GPTNeoForCausalLM, AutoConfig
+import torch
+from lm_train.utils import *
+import math
+
+class GPT:
+    def __init__(self, model_dtype="bf16", model_device="cuda"):
+        self.config = self.get_config(model_dtype, model_device)
+        self.checkpoint = self.get_checkpoint()
+        self.model = None
+        return
+    
+    def get_config(self, model_dtype="bf16", model_device="cuda"):
+        print("Using device:", model_device)
+        config = AutoConfig.from_pretrained("EleutherAI/gpt-neo-2.7B")
+        config.num_layers = 28
+        config.attention_layers = ["global"] * config.num_layers
+        config.attention_types = [["global"], config.num_layers]
+        config.num_heads = 16
+        config.hidden_size = 256 * config.num_heads
+        config.vocab_size = 50400
+        config.rotary = True
+        config.rotary_dim = 64
+        config.jax = True
+        config.model_dtype = model_dtype
+        config.model_device = model_device
+        if model_dtype == "bf16":
+            config.full_bf16 = True
+        return config
+
+    def get_checkpoint(self):
+        try:
+            from collections.abc import MutableMapping
+        except ImportError:
+            from collections import MutableMapping
+        from pathlib import Path
+
+        class Checkpoint(MutableMapping):
+            def __init__(self, chkpt_dir, device="cpu"):
+                self.device = device
+                self.chkpt_dir = Path(chkpt_dir)
+                self.checkpoint = torch.load(str(chkpt_dir / Path("m.pt")))
+            def __len__(self):
+                return len(self.checkpoint)
+            def __getitem__(self, key):
+                path = self.chkpt_dir / Path(self.checkpoint[key]).name
+                return torch.load(str(path), map_location=self.device)
+            def __setitem__(self, key, value):
+                return
+            def __delitem__(self, key, value):
+                return
+            def keys(self):
+                return self.checkpoint.keys()
+            def __iter__(self):
+                for key in self.checkpoint:
+                    yield (key, self.__getitem__(key))
+            def __copy__(self):
+                return Checkpoint(self.chkpt_dir, device=self.device)
+            def copy(self):
+                return Checkpoint(self.chkpt_dir, device=self.device)
+        
+        return Checkpoint
+        
+    def load_model(self, model_path=None, model_name=None, config=None, checkpoint=None):
+        if config == None:
+            config = self.config
+
+        if checkpoint == None:
+            Checkpoint = self.checkpoint
+
+        if model_name != None:
+            model_path = self.assign_path(model_name)
+            print("Loading model from: " + model_path)
+
+        model = no_init(lambda: GPTNeoForCausalLM.from_pretrained(pretrained_model_name_or_path=None, config=self.config, state_dict=Checkpoint(model_path)))
+        self.model = model
+        return model
+
+    def assign_path(self, model_name):
+        if model_name == "gptj":
+            return "/home/xuser/models/j6b_ckpt_14001"
+
+        # Raise error if model name not recognized
+        else:
+            raise ValueError("Model name not recognized")
+
+    def init_model(self, config=None, method='wang'):
+        neox_init = True
+        if config == None:
+            config = self.config
+
+        model = no_init(lambda: GPTNeoForCausalLM(config))
+        if neox_init:
+            modules = [*model.transformer.h[:-1], model.transformer.wte, model.transformer.ln_f]
+            init = small_init_method(self.config.hidden_size)
+            for module in modules:
+                for param in module.parameters():
+                    init(param)
+                    
+            last_layer = model.transformer.h[-1]
+            last_layer_init = wang_init_method(self.config.num_layers, self.config.hidden_size)
+            for param in last_layer.parameters():
+                last_layer_init(param)
+
+        self.model = model
+        return model
+
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        embs=None,
+    ):
+        if isinstance(self.model, GPTNeoForCausalLM):
+            outputs = self.model(input_ids, past_key_values, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict, embs)
+            # outputs: dict(loss, logits, past_key_values, hidden_states, attentions)
+        return outputs
+    
+#def init_module()
+
+def wang_init_method(n_layers, dim):
+    std = 2 / n_layers / math.sqrt(dim)
+
+    def init_(tensor):
+        return torch.nn.init.normal_(tensor, mean=0.0, std=std)
+
+    return init_
+
+# Stolen from NeoX. For the 20B run wang_init used on the output layer and small_init on rest of the layers.
+def small_init_method(dim):
+    """Fills the input Tensor with values according to the method described in Transformers without Tears: Improving 
+    the Normalization of Self-Attention - Nguyen, T. & Salazar, J. (2010), using a normal distribution."""
+    std = math.sqrt(2 / (5 * dim))
+
+    def init_(tensor):
+        return torch.nn.init.normal_(tensor, mean=0.0, std=std)
+
+    return init_
+
+

lm_train/optimizer.py

+from torch import optim
+import numpy as np
+#Based Optimizer
+
+def lr_schedule(step, warmup_steps, anneal_steps, lr, end_lr):
+    warmup_percent = np.clip(step, 0, warmup_steps) / warmup_steps
+    anneal_percent = np.clip(step - warmup_steps, 0, anneal_steps) / anneal_steps
+    #cosine schedule for annealing
+    return lr * warmup_percent - (lr - end_lr) * (1 - np.cos(np.pi * anneal_percent)) / 2
+
+
+class BasedOptimizer:
+    def __init__(self, parameters, config, optimizer):
+        self.lr = config["lr"]
+        self.end_lr = config["end_lr"] if "end_lr" in config else self.lr
+        self.warmup_steps =  config["warmup_steps"] if "warmup_steps" in config else 1
+        self.anneal_steps = config["anneal_steps"] if "anneal_steps" in config else 1
+        self.total_steps = config["total_steps"] if "total_steps" in config else None
+        self.weight_decay = config["weight_decay"] if "weight_decay" in config else 0
+        self.tokens = config["tokens"] if "tokens" in config else None
+        self.epochs = config["epochs"] if "epochs" in config else None
+        # tokens and epochs should not be here. calculate it somewhere else and find how many steps, then pass to the BasedOptimizer
+        self.beta1 = config["beta1"] if "beta1" in config else 0.9
+        self.beta2 = config["beta2"] if "beta2" in config else 0.95
+        self.eps = config["eps"] if "eps" in config else 1e-4
+        self.max_lr = False
+        self.curr_step = 0
+        self.curr_lr = 0
+            
+        if optimizer == "adamw":
+            self.optimizer = optim.AdamW(parameters, lr=0, weight_decay=self.weight_decay, betas=(self.beta1, self.beta2), eps=self.eps)
+        
+    def step(self):
+        self.optimizer.step()
+        self.curr_step = self.curr_step + 1
+        self.curr_lr = lr_schedule(self.curr_step, self.warmup_steps, self.anneal_steps, self.lr, self.end_lr)
+
+        if not self.max_lr:
+            if self.curr_lr == self.end_lr:
+                print("max lr reached.")
+                self.max_lr = True
+                
+            for paramx in self.optimizer.param_groups:
+                paramx['lr'] = self.curr_lr
+        
+    def zero_grad(self):
+        self.optimizer.zero_grad()
+        
+    def print_info(self):
+        print(f"end_lr: {str(self.end_lr)}")
+        print(f"warmup_steps: {str(self.warmup_steps)}")
+        print(f"total_steps: {str(self.total_steps)}")
+        print(f"weight_decay: {str(self.weight_decay)}")
+        print(f"step: {str(self.curr_step)}")
+        if self.curr_step != 0:
+            print(f"curr_lr: {str(self.get_current_lr())}")
\ No newline at end of file

lm_train/utils.py

+from torch.utils import data
+from transformers.modeling_utils import no_init_weights
+import numpy as np
+import torch
+
+# Does this work with other block_sizes? doesn't seem to.
+class FbDataset(data.Dataset):
+    def __init__(self, block_size, map_file, max_samples=None):
+        self.half_blocks = False
+        if block_size is not None and int(block_size) < 2048:
+            self.half_blocks = True
+        self.npz = np.memmap(map_file, mode="r", dtype="uint16").reshape((-1, 2048))
+        self.samples = self.npz.shape[0]
+        if self.half_blocks:
+            self.samples *= 2
+        if not max_samples is None:
+            self.samples = min(self.samples, int(max_samples))
+        self.skip = 0
+
+    def __len__(self):
+        return self.samples
+
+    def __getitem__(self, _id):
+        nth = _id + self.skip
+        offset = 0
+        length = 2048
+        if self.half_blocks:
+            nth = _id // 2
+            offset = 1024 * (_id % 2)
+            length = 1024
+        data = torch.tensor(self.npz[nth][offset:offset+length].astype(np.int64))
+        return (data, data)
+
+# Make loading models faster by not letting pytorch initialize the weights.
+# Usage: no_init(lambda: load_model(...))
+
+def no_init(loading_code):
+    def dummy(self):
+        return
+    
+    modules = [torch.nn.Linear, torch.nn.Embedding, torch.nn.LayerNorm]
+    original = {}
+    for mod in modules:
+        original[mod] = mod.reset_parameters
+        mod.reset_parameters = dummy
+    with no_init_weights():
+        result = loading_code()
+    for mod in modules:
+        mod.reset_parameters = original[mod]
+    
+    return result
+
+# Count the parameters of a given pytorch model.
+
+def count_parameters(model, only_trainable=False):
+    return sum(p.numel() for p in model.parameters() if p.requires_grad or not only_trainable)
\ No newline at end of file

main.py

 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from torch.utils.checkpoint import checkpoint as ck
 from einops import rearrange, repeat
 try:
     from collections.abc import MutableMapping
@@ -71,6 +72,7 @@ class SplitCheckpoint(MutableMapping):
 def get_logits(x, embedding):
     return embedding(x)
 
+@torch.jit.script
 def gelu_new(x):
     return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
 
@@ -84,7 +86,7 @@ def fixed_pos_embedding(dim=None, seq_len=None, x=None):
 def rotate_every_two(x):
     x1 = x[:, :, :, ::2]
     x2 = x[:, :, :, 1::2]
-    x = torch.stack((-x2, x1), axis=-1)
+    x = torch.stack((-x2, x1), dim=-1)
     return rearrange(x, '... d j -> ... (d j)')
 
 def apply_rotary_pos_emb(x, sincos, offset=0):
@@ -129,7 +131,7 @@ def _attn(query, key, value, causal_mask, masked_bias,
     if attention_mask is not None:
         attn_weights = attn_weights + attention_mask
 
-    attn_weights = nn.Softmax(dim=-1)(attn_weights)
+    attn_weights = F.softmax(attn_weights, dim=-1)
     attn_weights = attn_weights.to(value.dtype)
 
     attn_output = torch.matmul(attn_weights, value).to(value.dtype)
@@ -202,32 +204,41 @@ class SelfAttention(nn.Module):
         return x
 
 class FeedForward(nn.Module):
-    def __init__(self, dim=768, hidden_dim=768*4, activation=nn.GELU, device="cuda", dtype=torch.float16):
+    def __init__(self, dim=768, hidden_dim=768*4, activation=nn.GELU(), device="cuda", dtype=torch.float16):
         super(FeedForward, self).__init__()
         self.ff1 = nn.Linear(dim, hidden_dim, device=device, dtype=dtype)
         self.ff2 = nn.Linear(hidden_dim, dim, device=device, dtype=dtype)
         self.activation = activation
 
-    def forward(self, x):
+    def forward(self, x, act_ck=False):
         x = self.ff1(x)
+        if act_ck:
+            ck(self.activation, x)
+        else:
             x = self.activation(x)
         x = self.ff2(x)
         return x
 
 class GPTLayer(nn.Module):
-    def __init__(self, attn=SelfAttention, ff=FeedForward, hidden_dim=768, n_head=4, eps=1e-6, activation=nn.GELU, device="cuda", dtype=torch.float16):
+    def __init__(self, attn=SelfAttention, ff=FeedForward, hidden_dim=768, n_head=4, eps=1e-6, activation=nn.GELU(), device="cuda", dtype=torch.float16):
         super(GPTLayer, self).__init__()
         self.hidden_dim = hidden_dim
         self.ln_preattn = nn.LayerNorm(hidden_dim, eps=eps, device=device, dtype=dtype)
         self.ff = ff(dim=hidden_dim, hidden_dim=hidden_dim*4, activation=activation, device=device, dtype=dtype)
         self.attn = attn(hidden_dim=hidden_dim, n_head=n_head, device=device, dtype=dtype)
 
-    def forward(self, x, hypernetwork):
+    def forward(self, x, hypernetwork=None, act_ck=False):
         residual = x
+        if act_ck:
+            x = ck(self.ln_preattn, x)
+            attn_out = ck(self.attn, x)
+
+        else:
             x = self.ln_preattn(x)
             attn_out = self.attn(x)
-        ff_out = self.ff(x)
-        x = residual + ff_out + attn_out
+
+        ff_out = self.ff(x, act_ck)
+        x = residual + attn_out + ff_out
         if hypernetwork:
             hyper_out = hypernetwork(x)
             x = x + hyper_out
@@ -248,16 +259,15 @@ class GPTModel(nn.Module):
             #TODO: Decouple more, maybe even init everything here, not sure. Not modular enough yet.
             #TODO: Do we want to pass a config object everywhere? I don't exactly like that but passing a lot of variables is a bit ugly too.
 
-    def forward(self, x, hypernetwork=None):
+    def forward(self, x, hypernetwork=None, act_ck=False):
         x = self.vocab_embed(x)
         for layer in self.layers:
-            x = layer(x, hypernetwork)
-
+            x = layer(x, hypernetwork, act_ck)
         x = self.ln_final(x)
         return x
 
-    def get_logits(self, x):
-        x = self.forward(x)
+    def get_logits(self, x, act_ck=False):
+        x = self.forward(x, act_ck=act_ck)
         x = self.lm_head(x)
         return x.float()
     

test_pyfra.py

@@ -23,6 +23,7 @@ env1.sh('pip install /home/xuser/hugessd/pytorch/torch-1.10.1+cu113-cp38-cp38-li
 env1.sh('pip install einops numpy')
 env1.sh('pip install tqdm')
 env1.sh('pip install /home/xuser/diffusionstorage/workspace/finetune/pokepls/transformers-repo')
+env1.sh('pip3 install einops==0.4.1')
 with always_rerun():
     print(f"Running {sys.argv[1]}")
     path.sh(f'python3 {sys.argv[1]}')
\ No newline at end of file