id
int64
0
190k
prompt
stringlengths
21
13.4M
docstring
stringlengths
1
12k
10,154
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def piecewise_linear_func_ret_func(xs, ys, x): assert x >= xs[0] and x <= xs[-1] return np.interp(x, xs, ys) The provided code snippet includes necessary dependencies for implementing the `piecewise_linear_func` function. Write a Python function `def piecewise_linear_func(xs, ys)` to solve the following problem: Return a function created by linear inerpolation. Here is the function: def piecewise_linear_func(xs, ys): """Return a function created by linear inerpolation.""" indices = np.argsort(xs) xs = [xs[i] for i in indices] ys = [ys[i] for i in indices] # pad left and right k = 1e5 delta_x_left = xs[0] - xs[1] delta_y_left = ys[0] - ys[1] delta_x_right = xs[-1] - xs[-2] delta_y_right = ys[-1] - ys[-2] xs = [xs[0] + delta_x_left * k] + xs + [xs[-1] + delta_x_right * k] ys = [ys[0] + delta_y_left * k] + ys + [ys[-1] + delta_y_right * k] return functools.partial(piecewise_linear_func_ret_func, xs, ys)
Return a function created by linear inerpolation.
10,155
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def sample_from_range(n, k): assert n >= 1 if k == -1: ret = [1] while ret[-1] * 2 < n: ret.append(ret[-1] * 2) return ret else: if k == 1: return [1] step = (n - 1) // (k - 1) return list(range(1, n + 1, step))
null
10,156
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def cpu_mem_stats(): objects = gc.get_objects() tensors = [obj for obj in objects if torch.is_tensor(obj) and not obj.is_cuda] total_numel = 0 total_mem = 0 visited_data = set() for tensor in tensors: # a data_ptr indicates a memory block allocated data_ptr = tensor.storage().data_ptr() if data_ptr in visited_data: continue visited_data.add(data_ptr) numel = tensor.numel() total_numel += numel element_size = tensor.storage().element_size() mem = numel * element_size total_mem += mem return total_mem
null
10,157
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def torch_mem_stats(): objects = gc.get_objects() tensors = [obj for obj in objects if torch.is_tensor(obj) and obj.is_cuda] total_numel = 0 total_mem = 0 visited_data = set() for tensor in tensors: # a data_ptr indicates a memory block allocated data_ptr = tensor.storage().data_ptr() if data_ptr in visited_data: continue visited_data.add(data_ptr) print(tensor.shape, tensor.data_ptr()) numel = tensor.numel() total_numel += numel element_size = tensor.storage().element_size() mem = numel * element_size total_mem += mem return total_mem
null
10,158
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def array_1d(a, cls): return [cls() for _ in range(a)]
null
10,159
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def array_2d(a, b, cls): return [[cls() for _ in range(b)] for _ in range(a)]
null
10,160
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def array_3d(a, b, c, cls): return [[[cls() for _ in range(c)] for _ in range(b)] for _ in range(a)]
null
10,161
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch def array_4d(a, b, c, d, cls): return [[[[cls() for _ in range(d)] for _ in range(c)] for _ in range(b)] for _ in range(a)]
null
10,162
import argparse import dataclasses from attr import define, field from attr.setters import frozen import functools import gc import math import os from typing import Tuple, Union, Optional, Any, Sequence, List import numpy as np import torch class BenchmarkResult: def read_benchmark_log(filename): with open(filename) as fin: lines = fin.readlines() def extract(line): a, b = line.split("\t") latency = a[a.index(":") + 1:a.index(" s")] throughput = b[b.index(":") + 1:b.index(" to")] return float(latency), float(throughput) prefill_latency, prefill_throughput = extract(lines[2]) decode_latency, decode_throughput = extract(lines[3]) total_latency, total_throughput = extract(lines[4]) return BenchmarkResult( prefill_latency, prefill_throughput, decode_latency, decode_throughput, total_latency, total_throughput, )
null
10,163
import numpy as np import torch from flexgen.profile_bandwidth import benchmark_func def benchmark_func(func, number, repeat, warmup=3): for i in range(warmup): func() costs = [0] for i in range(repeat): torch.cuda.synchronize() tic = time.time() for i in range(number): func() torch.cuda.synchronize() costs.append((time.time() - tic) / number) return costs def bench_matmul(): for device in ["cuda", "cpu"]: for n in [1024, 2048]: if device == "cuda": dtype = torch.float16 else: dtype = torch.float32 a = torch.rand(n, n).to(dtype).to(device) b = torch.rand(n, n).to(dtype).to(device) def func(): return torch.matmul(a, b) cost = np.mean(benchmark_func(func, number=5, repeat=3)) tflops = 2 * n * n * n / cost / 1e12 print(f"device: {device}, N: {n}, latency: {cost*1e3:.2f} ms, TFLOPS: {tflops:.3f}") print()
null
10,164
import dataclasses import torch import numpy as np from flexgen.pytorch_backend import (TorchTensor, TorchDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.utils import np_dtype_to_torch_dtype default_cache_config = CompressionConfig( num_bits=0, group_size=0, group_dim=0, symmetric=False, enabled=False) def set_cache_compression_config(config): global default_cache_config default_cache_config = config
null
10,165
import dataclasses import torch import numpy as np from flexgen.pytorch_backend import (TorchTensor, TorchDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.utils import np_dtype_to_torch_dtype default_cache_config = CompressionConfig( num_bits=0, group_size=0, group_dim=0, symmetric=False, enabled=False) def get_cache_compression_config(): return default_cache_config
null
10,166
import dataclasses import torch import numpy as np from flexgen.pytorch_backend import (TorchTensor, TorchDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.utils import np_dtype_to_torch_dtype def compress(tensor, config): """Simulate group-wise quantization.""" if not config.enabled: return tensor group_size, num_bits, group_dim, symmetric = ( config.group_size, config.num_bits, config.group_dim, config.symmetric) assert num_bits <= 8 original_shape = tensor.shape num_groups = (original_shape[group_dim] + group_size - 1) // group_size new_shape = (original_shape[:group_dim] + (num_groups, group_size) + original_shape[group_dim+1:]) # Pad pad_len = (group_size - original_shape[group_dim] % group_size) % group_size if pad_len != 0: pad_shape = original_shape[:group_dim] + (pad_len,) + original_shape[group_dim+1:] tensor = torch.cat([ tensor, torch.zeros(pad_shape, dtype=tensor.dtype, device=tensor.device)], dim=group_dim) data = tensor.view(new_shape) # Quantize if symmetric: B = 2 ** (num_bits - 1) - 1 scale = B / torch.max(data.abs(), dim=group_dim + 1, keepdim=True)[0] data = data * scale data = data.clamp_(-B, B).round_().to(torch.int8) return data, scale, original_shape else: B = 2 ** num_bits - 1 mn = torch.min(data, dim=group_dim + 1, keepdim=True)[0] mx = torch.max(data, dim=group_dim + 1, keepdim=True)[0] scale = B / (mx - mn) data = data - mn data.mul_(scale) data = data.clamp_(0, B).round_().to(torch.uint8) return data, mn, scale, original_shape def decompress(packed_data, config): """Simulate group-wise dequantization.""" if not config.enabled: return packed_data group_size, num_bits, group_dim, symmetric = ( config.group_size, config.num_bits, config.group_dim, config.symmetric) # Dequantize if symmetric: data, scale, original_shape = packed_data data = data / scale else: data, mn, scale, original_shape = packed_data data = data / scale data.add_(mn) # Unpad pad_len = (group_size - original_shape[group_dim] % group_size) % group_size if pad_len: padded_original_shape = ( original_shape[:group_dim] + (original_shape[group_dim] + pad_len,) + original_shape[group_dim+1:]) data = data.reshape(padded_original_shape) indices = [slice(0, x) for x in original_shape] return data[indices].contiguous() else: return data.view(original_shape) def compress_and_decompress(tensor, config): packed_data = compress(tensor, config) return decompress(packed_data, config)
null
10,167
import dataclasses import torch import numpy as np from flexgen.pytorch_backend import (TorchTensor, TorchDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.utils import np_dtype_to_torch_dtype class CompressionConfig: """Group-wise quantization.""" num_bits: int group_size: int group_dim: int symmetric: bool enabled: bool = True def compress(tensor, config): """Simulate group-wise quantization.""" if not config.enabled: return tensor group_size, num_bits, group_dim, symmetric = ( config.group_size, config.num_bits, config.group_dim, config.symmetric) assert num_bits <= 8 original_shape = tensor.shape num_groups = (original_shape[group_dim] + group_size - 1) // group_size new_shape = (original_shape[:group_dim] + (num_groups, group_size) + original_shape[group_dim+1:]) # Pad pad_len = (group_size - original_shape[group_dim] % group_size) % group_size if pad_len != 0: pad_shape = original_shape[:group_dim] + (pad_len,) + original_shape[group_dim+1:] tensor = torch.cat([ tensor, torch.zeros(pad_shape, dtype=tensor.dtype, device=tensor.device)], dim=group_dim) data = tensor.view(new_shape) # Quantize if symmetric: B = 2 ** (num_bits - 1) - 1 scale = B / torch.max(data.abs(), dim=group_dim + 1, keepdim=True)[0] data = data * scale data = data.clamp_(-B, B).round_().to(torch.int8) return data, scale, original_shape else: B = 2 ** num_bits - 1 mn = torch.min(data, dim=group_dim + 1, keepdim=True)[0] mx = torch.max(data, dim=group_dim + 1, keepdim=True)[0] scale = B / (mx - mn) data = data - mn data.mul_(scale) data = data.clamp_(0, B).round_().to(torch.uint8) return data, mn, scale, original_shape def decompress(packed_data, config): """Simulate group-wise dequantization.""" if not config.enabled: return packed_data group_size, num_bits, group_dim, symmetric = ( config.group_size, config.num_bits, config.group_dim, config.symmetric) # Dequantize if symmetric: data, scale, original_shape = packed_data data = data / scale else: data, mn, scale, original_shape = packed_data data = data / scale data.add_(mn) # Unpad pad_len = (group_size - original_shape[group_dim] % group_size) % group_size if pad_len: padded_original_shape = ( original_shape[:group_dim] + (original_shape[group_dim] + pad_len,) + original_shape[group_dim+1:]) data = data.reshape(padded_original_shape) indices = [slice(0, x) for x in original_shape] return data[indices].contiguous() else: return data.view(original_shape) def test_simulated_compression(): torch.manual_seed(0) a = torch.normal(0, 1, (64, 64, 64), dtype=torch.float16).cuda() config = CompressionConfig( num_bits=4, group_size=32, group_dim=0, symmetric=False) packed_data = compress(a, config) b = decompress(packed_data, config) print(a[0]) print(b[0])
null
10,168
import dataclasses import torch import numpy as np from flexgen.pytorch_backend import (TorchTensor, TorchDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.utils import np_dtype_to_torch_dtype class CompressionConfig: """Group-wise quantization.""" num_bits: int group_size: int group_dim: int symmetric: bool enabled: bool = True def compress(tensor, config): """Simulate group-wise quantization.""" if not config.enabled: return tensor group_size, num_bits, group_dim, symmetric = ( config.group_size, config.num_bits, config.group_dim, config.symmetric) assert num_bits <= 8 original_shape = tensor.shape num_groups = (original_shape[group_dim] + group_size - 1) // group_size new_shape = (original_shape[:group_dim] + (num_groups, group_size) + original_shape[group_dim+1:]) # Pad pad_len = (group_size - original_shape[group_dim] % group_size) % group_size if pad_len != 0: pad_shape = original_shape[:group_dim] + (pad_len,) + original_shape[group_dim+1:] tensor = torch.cat([ tensor, torch.zeros(pad_shape, dtype=tensor.dtype, device=tensor.device)], dim=group_dim) data = tensor.view(new_shape) # Quantize if symmetric: B = 2 ** (num_bits - 1) - 1 scale = B / torch.max(data.abs(), dim=group_dim + 1, keepdim=True)[0] data = data * scale data = data.clamp_(-B, B).round_().to(torch.int8) return data, scale, original_shape else: B = 2 ** num_bits - 1 mn = torch.min(data, dim=group_dim + 1, keepdim=True)[0] mx = torch.max(data, dim=group_dim + 1, keepdim=True)[0] scale = B / (mx - mn) data = data - mn data.mul_(scale) data = data.clamp_(0, B).round_().to(torch.uint8) return data, mn, scale, original_shape def decompress(packed_data, config): """Simulate group-wise dequantization.""" if not config.enabled: return packed_data group_size, num_bits, group_dim, symmetric = ( config.group_size, config.num_bits, config.group_dim, config.symmetric) # Dequantize if symmetric: data, scale, original_shape = packed_data data = data / scale else: data, mn, scale, original_shape = packed_data data = data / scale data.add_(mn) # Unpad pad_len = (group_size - original_shape[group_dim] % group_size) % group_size if pad_len: padded_original_shape = ( original_shape[:group_dim] + (original_shape[group_dim] + pad_len,) + original_shape[group_dim+1:]) data = data.reshape(padded_original_shape) indices = [slice(0, x) for x in original_shape] return data[indices].contiguous() else: return data.view(original_shape) class TorchDevice: """Wrap tensor and computation APIs of a single CPU or GPU.""" def __init__(self, name, mem_capacity=None, flops=None): self.name = name self.mem_capacity = mem_capacity self.flops = flops self.dev = torch.device(name) self.device_type = DeviceType.convert(self.dev.type) self.compressed_device = TorchCompressedDevice(self) self.links = {} self.attention_compute_workspace = None self.workspace_pt = 0 if self.device_type == DeviceType.CPU: global global_cpu_device global_cpu_device = self def add_link(self, link): dst = link.b if link.a == self else link.a self.links[dst] = link def allocate(self, shape, dtype, pin_memory=None, name=None): if self.device_type == DeviceType.CPU: pin_memory = True if pin_memory is None else pin_memory else: pin_memory = False dtype = np_dtype_to_torch_dtype[dtype] data = torch.empty(shape, dtype=dtype, pin_memory=pin_memory, device=self.dev) return TorchTensor.create_from_torch(data, self, name=name) def delete(self, tensor): pass def init_attention_compute_workspace(self, config, task, policy): if self.device_type != DeviceType.CPU: return # Only CPU requires this fp32 workspace if not policy.compress_cache: b = policy.gpu_batch_size n_head = config.n_head head_dim = config.input_dim // n_head max_seq_len = task.prompt_len + task.gen_len - 1 self.attention_compute_workspace = [] self.workspace_pt = 0 # We currently separate SelfAttention and MLP as two layers, # so we only need one workspace instead of two. for i in range(1 if policy.sep_layer else 2): shape = (max_seq_len, b * n_head, head_dim) k_cache = self.allocate(shape, np.float32, pin_memory=False) v_cache = self.allocate(shape, np.float32, pin_memory=False) self.attention_compute_workspace.append((k_cache, v_cache)) else: self.compressed_device.init_attention_compute_workspace( config, task, policy) def next_attention_compute_workspace(self): self.workspace_pt = (self.workspace_pt + 1) % len( self.attention_compute_workspace) return self.attention_compute_workspace[self.workspace_pt] def del_attention_compute_workspace(self): self.attention_compute_workspace = None def gen_attention_mask(self, token_ids, pad_token_id, donate): data = token_ids.data.ne(pad_token_id) if donate[0]: token_ids.delete() return TorchTensor.create_from_torch(data, self) def extend_attention_mask(self, attention_mask, donate): bs = attention_mask.shape[0] data = torch.concat((attention_mask.data, torch.ones((bs, 1), dtype=attention_mask.dtype, device=self.dev)), dim=1) if donate[0]: attention_mask.delete() return TorchTensor.create_from_torch(data, self) def opt_input_embed(self, inputs, attention_mask, w_token, w_pos, pad_token_id, donate): # decompress weights if w_token.device.device_type == DeviceType.COMPRESSED: w_token = w_token.device.decompress(w_token) w_pos = w_pos.device.decompress(w_pos) token_ids = inputs.data mask = attention_mask.data if donate[0]: inputs.delete() if donate[1]: attention_mask.delete() # token embedding token_embed = F.embedding(token_ids, w_token.data, pad_token_id) # pos embedding positions = torch.cumsum(mask, dim=1).int() * mask + 1 # cut positions if `past_key_values_length` is > 0 past_key_values_length = mask.shape[1] - token_ids.shape[1] positions = positions[:, past_key_values_length:] pos_embed = F.embedding(positions, w_pos.data) data = token_embed + pos_embed return TorchTensor.create_from_torch(data, self) def opt_output_embed(self, inputs, w_ln, b_ln, w_token, donate, do_sample, temperature): # decompress weights if w_token.device.device_type == DeviceType.COMPRESSED: w_token = w_token.device.decompress(w_token) b, s, h = inputs.shape hidden = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) if donate[0]: inputs.delete() # output embedding logits = F.linear(hidden, w_token.data) last_token_logits = logits[:,-1,:] if do_sample and not temperature < 1e-5: probs = torch.softmax(last_token_logits / temperature, dim=-1) ids = torch.multinomial(probs, num_samples=1) else: ids = last_token_logits.argmax(dim=1, keepdim=True) return TorchTensor.create_from_torch(ids, self) def init_cache_one_gpu_batch(self, config, task, policy): num_head, hidden_size, prompt_len, gen_len, gpu_batch_size = ( config.n_head, config.input_dim, task.prompt_len, task.gen_len, policy.gpu_batch_size) shape = (prompt_len + gen_len - 1, gpu_batch_size * num_head, hidden_size // num_head) # NOTE: disable pin_memory due to high memory overhead pin_memory = False k_cache = self.allocate(shape, np.float16, pin_memory=pin_memory) v_cache = self.allocate(shape, np.float16, pin_memory=pin_memory) return k_cache, v_cache def mha(self, inputs, attention_mask, w_q, b_q, w_k, b_k, w_v, b_v, w_out, b_out, w_ln, b_ln, n_head, donate, compress_cache, comp_config): """Multi-head attention (prefill phase).""" # decompress weights if w_q.device.device_type == DeviceType.COMPRESSED: w_q = w_q.device.decompress(w_q) w_k = w_k.device.decompress(w_k) w_v = w_v.device.decompress(w_v) w_out = w_out.device.decompress(w_out) b, s, h = inputs.shape head_dim = h // n_head scaling = head_dim ** -0.5 hidden = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) # shape: (b, s, h) q = F.linear(hidden, w_q.data, bias=b_q.data) * scaling k = F.linear(hidden, w_k.data, bias=b_k.data) v = F.linear(hidden, w_v.data, bias=b_v.data) # shape: (b, s, n_head, head_dim) q = q.view(b, s, n_head, head_dim) k = k.view(b, s, n_head, head_dim) v = v.view(b, s, n_head, head_dim) # shape: (b * n_head, s, head_dim) q = q.permute(0, 2, 1, 3).reshape(b * n_head, s, head_dim) # shape: (b * n_head, head_dim, s) k = k.permute(0, 2, 3, 1).reshape(b * n_head, head_dim, s) # shape: (b * n_head, s, head_dim) v = v.permute(0, 2, 1, 3).reshape(b * n_head, s, head_dim) # shape: (b * n_head, s, s) attn_weights = torch.bmm(q, k) # shape: (b, 1, s, s) idx = torch.arange(s, device=self.dev) causal_mask = (idx <= idx.view(s, 1)).view(1, 1, s, s) mask = attention_mask.data.view(b, 1, 1, s) & causal_mask # shape: (b, n_head, s, s) attn_weights = attn_weights.view(b, n_head, s, s) attn_weights = torch.where(mask, attn_weights, -1e4) attn_weights = attn_weights.view(b * n_head, s, s) attn_weights = F.softmax(attn_weights, dim=2) # shape: (b, n_head, s, head_dim) value = torch.bmm(attn_weights, v).view(b, n_head, s, head_dim) # shape: (b, s, h) value = value.transpose(1, 2).reshape(b, s, h) value = F.linear(value, w_out.data, bias=b_out.data) value.add_(inputs.data) if donate[0]: inputs.delete() if donate[1]: attention_mask.delete() # (s, b * n_head, head_dim) k = k.permute(2, 0, 1) v = v.permute(1, 0, 2) if compress_cache: k = self.compressed_device.compress(k, comp_config) v = self.compressed_device.compress(v, comp_config) else: k = TorchTensor.create_from_torch(k, self) v = TorchTensor.create_from_torch(v, self) return TorchTensor.create_from_torch(value, self), k, v def mha_gen(self, inputs, attention_mask, w_q, b_q, w_k, b_k, w_v, b_v, w_out, b_out, w_ln, b_ln, n_head, k_cache, v_cache, donate, attn_sparsity, compress_cache, comp_config): """Multi-head attention (decoding phase).""" # decompress weights if w_q.device.device_type == DeviceType.COMPRESSED: w_q = w_q.device.decompress(w_q) w_k = w_k.device.decompress(w_k) w_v = w_v.device.decompress(w_v) w_out = w_out.device.decompress(w_out) b, tgt_s, h = inputs.shape src_s = attention_mask.shape[1] head_dim = h // n_head scaling = head_dim ** -0.5 hidden = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) # shape: (b, 1, h) q = F.linear(hidden, w_q.data, bias=b_q.data) * scaling k = F.linear(hidden, w_k.data, bias=b_k.data) v = F.linear(hidden, w_v.data, bias=b_v.data) # shape: (b, 1, n_head, head_dim) q = q.view(b, tgt_s, n_head, head_dim) k = k.view(b, tgt_s, n_head, head_dim) v = v.view(b, tgt_s, n_head, head_dim) # shape: (b * n_head, 1, head_dim) q = q.permute(0, 2, 1, 3).reshape(b * n_head, tgt_s, head_dim) # shape: (1, b * n_head, head_dim) k_new = k.permute(1, 0, 2, 3).reshape(tgt_s, b * n_head, head_dim) # shape: (1, b * n_head, head_dim) v_new = v.permute(1, 0, 2, 3).reshape(tgt_s, b * n_head, head_dim) if isinstance(k_cache, TorchTensor): if attn_sparsity >= 1.0: # Dense attention if compress_cache: # shape: (s, b * n_head, head_dim) k = k_cache.device.decompress(k_cache)[:src_s] v = v_cache.device.decompress(v_cache)[:src_s] else: # shape: (s, b * n_head, head_dim) k = k_cache.data[:src_s] v = v_cache.data[:src_s] k[src_s - 1:src_s] = k_new v[src_s - 1:src_s] = v_new # shape: (b * n_head, head_dim, s) k = k.permute(1, 2, 0).reshape(b * n_head, head_dim, src_s) # shape: (b * n_head, s, head_dim) v = v.permute(1, 0, 2).reshape(b * n_head, src_s, head_dim) if k.is_cuda: value = self._attention_value(q, k, v, attention_mask.data, b, src_s, tgt_s, n_head, head_dim) else: q = q.float().cpu() k, v = k.float(), v.float() value = self._attention_value(q, k, v, attention_mask.data, b, src_s, tgt_s, n_head, head_dim).cuda().half() else: # Sparse attention # shape: (s, b * n_head, head_dim) k = k_cache.data[:src_s] k[src_s - 1:src_s] = k_new # shape: (b * n_head, head_dim, s) k = k.permute(1, 2, 0).reshape(b * n_head, head_dim, src_s) if k.is_cuda: value = self._sparse_attention_value(q, k, v_new, v_cache, attention_mask.data, b, src_s, tgt_s, n_head, head_dim, attn_sparsity) else: q = q.float().cpu() value = self._sparse_attention_value(q, k, v_new, v_cache, attention_mask.data, b, src_s, tgt_s, n_head, head_dim, attn_sparsity).cuda().half() else: # Mixed device attention assert attn_sparsity >= 1.0 value = self._mixed_device_attention(q, k_cache, v_cache, k_new, v_new, attention_mask.data, b, src_s, tgt_s, n_head, head_dim) # shape: (b, 1, h) value = value.transpose(1, 2).view(b, tgt_s, h) value = F.linear(value, w_out.data, bias=b_out.data) value.add_(inputs.data) if donate[0]: inputs.delete() if donate[1]: attention_mask.delete() if compress_cache: if comp_config.group_dim == 0: s_ = src_s // comp_config.group_size * comp_config.group_size k_new = k[:, :, s_:].permute(2, 0, 1) v_new = v[:, s_:, :].permute(1, 0, 2) k_new = self.compressed_device.compress(k_new, comp_config) v_new = self.compressed_device.compress(v_new, comp_config) else: k_new = TorchTensor.create_from_torch(k_new, self) v_new = TorchTensor.create_from_torch(v_new, self) return TorchTensor.create_from_torch(value, self), k_new, v_new def _attention_weights(self, q, k, mask, b, src_s, n_head): # shape: (b * n_head, 1, s) attn_weights = torch.bmm(q, k) # shape: (b, 1, 1, s) mask = mask.view(b, 1, 1, src_s) # shape: (b * n_head, 1, s) attn_weights = attn_weights.view(b, n_head, 1, src_s) attn_weights = torch.where(mask, attn_weights, -1e4) attn_weights = attn_weights.view(b * n_head, 1, src_s) attn_weights = F.softmax(attn_weights, dim=2) return attn_weights def _attention_value(self, q, k, v, mask, b, src_s, tgt_s, n_head, head_dim): # shape: (b * n_head, 1, s) attn_weights = self._attention_weights(q, k, mask, b, src_s, n_head) # shape: (b, n_head, 1, head_dim) return torch.bmm(attn_weights, v).view(b, n_head, tgt_s, head_dim) def _sparse_attention_value(self, q, k, v_new, v_cache, mask, b, src_s, tgt_s, n_head, head_dim, attn_sparsity): # shape: (b * n_head, 1, s) attn_weights = self._attention_weights(q, k, mask, b, src_s, n_head) topk = int(attn_sparsity * (attn_weights.shape[2] - 1)) topk_weights, topk_indices = attn_weights[:, :, :-1].topk( topk, dim=2, sorted=False) topk_indices = topk_indices.view(b * n_head, topk).transpose(0, 1) # shape: (b * n_head, 1, topk+1) attn_weights = torch.cat([topk_weights, attn_weights[:, :, -1].unsqueeze(-1)], dim=-1) if k.is_cuda: v_home = v_cache v_buf = self.allocate((topk+1, b*n_head, head_dim), np.float16) topk_indices = topk_indices.cpu() else: (v_home, v_buf) = v_cache # shape: (s, b * n_head, head_dim) indices_src = topk_indices indices_tgt = (slice(0, indices_src.shape[0]), slice(0, v_home.shape[1])) general_copy(v_buf, indices_tgt, v_home, indices_src) v_home.device.synchronize() # shape: (topk+1, b * n_head, head_dim) v = v_buf.data[:topk+1] v[topk:topk+1] = v_new # shape: (b * n_head, topk+1, head_dim) v = v.permute(1, 0, 2).reshape(b * n_head, topk+1, head_dim) # shape: (b * n_head, 1, head_dim) return torch.bmm(attn_weights, v).view(b, n_head, tgt_s, head_dim) def _mixed_device_attention(self, q, k_cache, v_cache, k_new, v_new, mask, b, src_s, tgt_s, n_head, head_dim): # The caches are stored on both gpu and cpu. # Compute attention on gpu for caches stored on gpu. # Compute attention on cpu for caches stored on cpu. k_gpu, k_cpu = k_cache[0].data, k_cache[1].data v_gpu, v_cpu = v_cache[0].data, v_cache[1].data seg = k_gpu.shape[1] # Compute GPU part b_gpu = seg // n_head q_gpu = q[:seg] # shape: (s, b * n_head, head_dim) k_gpu = k_gpu[:src_s, :seg, :] v_gpu = v_gpu[:src_s, :seg, :] k_gpu[src_s-1:src_s, :, :] = k_new[:, :seg, :] v_gpu[src_s-1:src_s, :, :] = v_new[:, :seg, :] # shape: (b * n_head, head_dim, s) k_gpu = k_gpu.permute(1, 2, 0) # shape: (b * n_head, s, head_dim) v_gpu = v_gpu.permute(1, 0, 2) mask_gpu = mask[:b_gpu].cuda() value_gpu = self._attention_value(q_gpu, k_gpu, v_gpu, mask_gpu, b_gpu, src_s, tgt_s, n_head, head_dim) # Compute CPU Part b_cpu = b - b_gpu q_cpu = q[seg:].float().cpu() # shape: (s, b * n_head, head_dim) k_cpu = k_cpu[:src_s, seg:, :] v_cpu = v_cpu[:src_s, seg:, :] k_cpu[src_s-1:src_s, :, :] = k_new[:, seg:, :] v_cpu[src_s-1:src_s, :, :] = v_new[:, seg:, :] # shape: (b * n_head, head_dim, s) k_cpu = k_cpu.permute(1, 2, 0) # shape: (b * n_head, s, head_dim) v_cpu = v_cpu.permute(1, 0, 2) mask_cpu = mask[b_gpu:] value_cpu = self._attention_value(q_cpu, k_cpu, v_cpu, mask_cpu, b_cpu, src_s, tgt_s, n_head, head_dim) value = torch.cat([value_gpu, value_cpu.cuda().half()], dim=0) return value def mlp(self, inputs, wi, bi, wo, bo, w_ln, b_ln, donate): # decompress weights if wi.device.device_type == DeviceType.COMPRESSED: wi = wi.device.decompress(wi) wo = wo.device.decompress(wo) b, s, h = inputs.shape out = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) out = F.linear(out, wi.data, bias=bi.data) F.relu(out, inplace=True) out = F.linear(out, wo.data, bias=bo.data) out.add_(inputs.data) if donate[0]: inputs.delete() return TorchTensor.create_from_torch(out, self) def synchronize(self): torch.cuda.synchronize() def mem_stats(self): if self.device_type == DeviceType.CUDA: cur_mem = torch.cuda.memory_allocated(self.dev) peak_mem = torch.cuda.max_memory_allocated(self.dev) elif self.device_type == DeviceType.CPU: cur_mem = cpu_mem_stats() peak_mem = 0 else: raise NotImplementedError() return cur_mem, peak_mem def print_stats(self, output_file=None): torch.cuda.synchronize() cur_mem, peak_mem = self.mem_stats() if output_file is not None: with open(output_file, "w") as f: f.write(f"TorchDevice: {self.name}\n") f.write(f" cur_mem: {cur_mem/GB:.4f} GB, " f" peak_mem: {peak_mem/GB:.4f} GB\n") else: print(f"TorchDevice: {self.name}") print(f" cur_mem: {cur_mem/GB:.4f} GB, " f" peak_mem: {peak_mem/GB:.4f} GB") return cur_mem, peak_mem def __str__(self): return f"TorchDevice(name={self.name})" def test_real_compression(): torch.manual_seed(0) a = torch.normal(0, 1, (32, 1, 1), dtype=torch.float16).cuda() config = CompressionConfig( num_bits=4, group_size=32, group_dim=0, symmetric=False) dev = TorchDevice("cuda:0", 0, 0).compressed_device packed = dev.compress(a, config) b = dev.decompress(packed) print(a.flatten()) print(b.flatten())
null
10,169
from enum import Enum, auto from functools import partial from itertools import count import os import queue import shutil import time import threading from typing import Optional, Union, Tuple import torch import torch.nn.functional as F import numpy as np from flexgen.utils import (GB, T, cpu_mem_stats, vector_gather, np_dtype_to_torch_dtype, torch_dtype_to_np_dtype, torch_dtype_to_num_bytes) general_copy_compressed = TorchCompressedDevice = None global_cpu_device = None def fix_recursive_import(): global general_copy_compressed, TorchCompressedDevice, global_cpu_device from flexgen import compression general_copy_compressed = compression.general_copy_compressed TorchCompressedDevice = compression.TorchCompressedDevice
null
10,170
from enum import Enum, auto from functools import partial from itertools import count import os import queue import shutil import time import threading from typing import Optional, Union, Tuple import torch import torch.nn.functional as F import numpy as np from flexgen.utils import (GB, T, cpu_mem_stats, vector_gather, np_dtype_to_torch_dtype, torch_dtype_to_np_dtype, torch_dtype_to_num_bytes) class DeviceType(Enum): CPU = auto() CUDA = auto() DISK = auto() MIXED = auto() COMPRESSED = auto() def convert(name): if name == "cpu": return DeviceType.CPU elif name == "cuda": return DeviceType.CUDA elif name == "disk": return DeviceType.DISK elif name == "mixed": return DeviceType.MIXED elif name == "compressed": return DeviceType.COMPRESSED else: raise ValueError(f"Invalid name: {name}") def map_to_torch_tensor(tensor, indices): if tensor.device.device_type == DeviceType.DISK: data = torch.from_numpy(np.lib.format.open_memmap(tensor.data)) else: data = tensor.data # BC: this is supposed to only handle the sparse v_cache case if torch.is_tensor(indices): return vector_gather(data, indices) return data[indices] if indices else data GB = 1 << 30 The provided code snippet includes necessary dependencies for implementing the `copy_worker_func` function. Write a Python function `def copy_worker_func(queue, cuda_id)` to solve the following problem: The copy worker thread. Here is the function: def copy_worker_func(queue, cuda_id): """The copy worker thread.""" torch.cuda.set_device(cuda_id) cpu_buf = torch.empty((1 * GB,), dtype=torch.float16, pin_memory=True) copy_stream = torch.cuda.Stream() with torch.cuda.stream(copy_stream): while True: item = queue.get() if item is None: queue.task_done() return dst, dst_indices, src, src_indices = item src_data = map_to_torch_tensor(src, src_indices) dst_data = map_to_torch_tensor(dst, dst_indices) if (src.device.device_type == DeviceType.CUDA or dst.device.device_type == DeviceType.CUDA): # Use a pinned cpu buffer as a relay size = np.prod(src_data.shape) tmp_cpu_buf = cpu_buf[:size].view(src_data.shape) tmp_cpu_buf.copy_(src_data) dst_data.copy_(tmp_cpu_buf) else: dst_data.copy_(src_data) queue.task_done()
The copy worker thread.
10,171
import argparse from dataclasses import asdict, replace import json import math import os import time from flexgen.flex_opt import (Policy, OptLM, ExecutionEnv, CompressionConfig, str2bool) from helm.benchmark.presentation.run_entry import RunEntry from helm.benchmark.run import run_entries_to_run_specs from helm.benchmark.run_specs import (ScenarioSpec, RunSpec, get_summarization_adapter_spec, get_summarization_metric_specs, get_generative_harms_metric_specs, ADAPT_MULTIPLE_CHOICE_JOINT, get_multiple_choice_adapter_spec) from helm.benchmark.runner import (create_scenario, AdapterFactory, with_instance_ids, create_metric, TokensMetric, Metric, MetricSpec, MetricResult, PerInstanceStats, create_metric, Stat, ScenarioState, Counter, MetricName, ensure_directory_exists, write, asdict_without_nones, DataPreprocessor) from helm.common.request import Request, RequestResult, Sequence, Token from helm.common.tokenization_request import (TokenizationRequestResult, TokenizationRequest, TokenizationToken, DecodeRequest, DecodeRequestResult) from helm.proxy.clients.client import truncate_sequence import numpy as np from tqdm import tqdm from transformers import AutoTokenizer, AutoConfig class OptTokenizer: # Adapted from helm/proxy/clients/huggingface_client.py def __init__(self, name): self.tokenizer = AutoTokenizer.from_pretrained(name, padding_side="left") self.tokenizer.add_bos_token = False if 'galactica' in name: config = AutoConfig.from_pretrained(name) self.tokenizer.pad_token = config.pad_token_id self.tokenizer.eos_token = config.eos_token_id def tokenize(self, request: TokenizationRequest) -> TokenizationRequestResult: tokenizer = self.tokenizer def do_it(): if request.encode: if request.truncation: tokens = tokenizer.encode( request.text, truncation=request.truncation, max_length=request.max_length, add_special_tokens=False, ) else: tokens = tokenizer.encode(request.text, add_special_tokens=False) else: tokens = tokenizer.tokenize(request.text) return {"tokens": tokens} result = do_it() return TokenizationRequestResult( success=True, cached=False, text=request.text, tokens=[TokenizationToken(value) for value in result["tokens"]], request_time=0, ) def decode(self, request: DecodeRequest) -> DecodeRequestResult: tokenizer = self.tokenizer def do_it(): return { "text": tokenizer.decode( request.tokens, clean_up_tokenization_spaces=request.clean_up_tokenization_spaces ) } result = do_it() return DecodeRequestResult( success=True, cached=False, text=result["text"], request_time=0, ) def execute(scenario_state, tokenizer, effective_bs, pad_to_seq_len): generation_args = get_hf_generation_args( scenario_state.request_states[0].request, tokenizer) batches = get_batches(scenario_state, tokenizer, effective_bs, pad_to_seq_len=pad_to_seq_len) # Initialize environment env = ExecutionEnv.create(args.offload_dir) # Offloading policy policy = Policy(args.gpu_batch_size, args.num_gpu_batches, args.percent[0], args.percent[1], args.percent[2], args.percent[3], args.percent[4], args.percent[5], overlap=True, sep_layer=True, pin_weight=args.pin_weight, cpu_cache_compute=args.cpu_cache_compute, attn_sparsity=1.0, compress_weight=args.compress_weight, comp_weight_config=CompressionConfig( num_bits=4, group_size=64, group_dim=0, symmetric=False), compress_cache=args.compress_cache, comp_cache_config=CompressionConfig( num_bits=4, group_size=64, group_dim=2, symmetric=False)) print(f"Init weights begin.") tic = time.time() model = OptLM(args.model, env, args.path, policy) print(f"Init weights end. Elapsed: {time.time() - tic:.2f} s", flush=True) # Generate print(f"Generate begin. #sequences: {len(batches) * effective_bs}") tic = time.time() input_ids_batches = [] output_ids_batches = [] for batch in tqdm(batches): input_ids_tmp = batch["input_ids"] output_ids_tmp = model.generate( input_ids_tmp, do_sample=generation_args["do_sample"], temperature=generation_args["temperature"], max_new_tokens=generation_args["max_new_tokens"], stop=generation_args.get("eos_token_id", None)) input_ids_batches.append(input_ids_tmp) output_ids_batches.append(output_ids_tmp) print(f"Generate end. Elapsed: {time.time() - tic:.2f} s", flush=True) input_ids = np.concatenate(input_ids_batches) output_ids = np.concatenate(output_ids_batches) outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True) #print("Outputs:\n" + 70 * '-') ##for i in range(len(outputs)): #for i in [0, len(outputs) - 1]: # print(f"{i}:\n{outputs[i]}") # print("-" * 70) # Shutdown print("Shutdown...") env.close_copy_threads() request_states = [] for i, request_state in enumerate(scenario_state.request_states): request = request_state.request encoded_input = input_ids[i] sequences = [output_ids[i]] if not request.echo_prompt: sequences = [sequence[len(encoded_input) :] for sequence in sequences] all_tokens = [tokenizer.convert_ids_to_tokens(sequence) for sequence in sequences] all_decoded_text = tokenizer.batch_decode(sequences) all_logprobs_of_chosen_tokens = [[0] * len(x) for x in all_tokens] all_top_logprobs_dicts = [[{}] * len(x) for x in all_tokens] completions = [] for (decoded_text, tokens, logprobs_of_chosen_tokens, top_logprobs_dicts) in zip( all_decoded_text, all_tokens, all_logprobs_of_chosen_tokens, all_top_logprobs_dicts ): completions.append( { "text": decoded_text, "tokens": tokens, "logprobs": logprobs_of_chosen_tokens, "top_logprobs_dicts": top_logprobs_dicts, } ) response = { "completions": completions, "input_length": len(encoded_input)} completions = [] for raw_completion in response["completions"]: sequence_logprob: float = 0 tokens: List[Token] = [] if request.echo_prompt: # Add prompt to list of generated tokens. generated_tokens = raw_completion["tokens"][response["input_length"] :] for token_text in raw_completion["tokens"][: response["input_length"]]: tokens.append(Token(text=token_text, logprob=0.0, top_logprobs={})) else: generated_tokens = raw_completion["tokens"] # Compute logprob for the entire sequence. for token_text, logprob, top_logprobs_dict in zip( generated_tokens, raw_completion["logprobs"], raw_completion["top_logprobs_dicts"] ): tokens.append(Token(text=token_text, logprob=logprob, top_logprobs=top_logprobs_dict)) sequence_logprob += logprob completion = Sequence(text=raw_completion["text"], logprob=sequence_logprob, tokens=tokens) completion = truncate_sequence(completion, request) completions.append(completion) result = RequestResult( success=True, cached=False, request_time=0, request_datetime=0, completions=completions, embedding=[], ) request_states.append(replace(request_state, result=result)) return ScenarioState(scenario_state.adapter_spec, request_states) def run_entry(description, pad_to_seq_len, args): effective_bs = args.gpu_batch_size * args.num_gpu_batches parallelism = 4 ##### RunSpec ##### run_entries = [RunEntry(description, priority=1, groups=None)] run_specs = run_entries_to_run_specs( run_entries=run_entries, max_eval_instances=args.max_eval_instances, num_train_trials=3, ) run_spec = run_specs[0] run_path: str = os.path.join(args.run_path, run_spec.name) ensure_directory_exists(run_path) eval_cache_path: str = os.path.join(run_path, "eval_cache") ensure_directory_exists(eval_cache_path) ##### Adapter ##### #tokenizer_service = OptTokenizer("facebook/opt-30b") tokenizer_service = OptTokenizer(args.model) tokenizer = tokenizer_service.tokenizer adapter = AdapterFactory.get_adapter(run_spec.adapter_spec, tokenizer_service) ##### Scenario ##### print(run_spec) scenario = create_scenario(run_spec.scenario_spec) scenario.output_path = f"data/{run_spec.name}" os.makedirs(scenario.output_path, exist_ok=True) instances = scenario.get_instances() # Give each instance a unique ID instances = with_instance_ids(instances) # Get the instances necessary for this run. instances = adapter.get_run_instances(instances) # Data preprocessing instances = DataPreprocessor(run_spec.data_augmenter_spec).preprocess( instances, parallelism=parallelism ) scenario_state = adapter.adapt(instances, parallelism=parallelism) ##### Execute ##### if pad_to_seq_len is None: pad_to_seq_len = adapter.window_service.max_sequence_length - run_spec.adapter_spec.max_tokens + 1 scenario_state = execute(scenario_state, tokenizer, effective_bs, pad_to_seq_len) ##### Metrics ##### metrics = ( [create_metric(metric_spec) for metric_spec in run_spec.metric_specs]) + [TokensMetric()] metrics = [metrics[0]] stats: List[Stat] = [] per_instance_stats: List[PerInstanceStats] = [] for metric in metrics: metric_result: MetricResult = metric.evaluate( scenario_state, tokenizer_service, eval_cache_path, parallelism=parallelism, ) stats.extend(metric_result.aggregated_stats) per_instance_stats.extend(metric_result.per_instance_stats) # Check that there aren't duplicate `Stat`s # Note: doesn't catch near misses. metric_counts: typing.Counter[MetricName] = Counter([stat.name for stat in stats]) for metric_name, count in metric_counts.items(): if count > 1: print(f"WARNING: duplicate metric name {metric_name}") # Print out the number of stats print(f"Generated {len(stats)} stats.") # Output benchmarking information and results to files write(os.path.join(run_path, "run_spec.json"), json.dumps(asdict_without_nones(run_spec), indent=2)) # Write out scenario write(os.path.join(run_path, "scenario.json"), json.dumps(asdict_without_nones(scenario), indent=2)) # Write scenario state write(os.path.join(run_path, "scenario_state.json"), json.dumps(asdict_without_nones(scenario_state), indent=2)) write( os.path.join(run_path, "stats.json"), json.dumps([asdict_without_nones(stat) for stat in stats], indent=2) ) write( os.path.join(run_path, "per_instance_stats.json"), json.dumps(list(map(asdict_without_nones, per_instance_stats)), indent=2), )
null
10,172
import argparse from tqdm import tqdm import json import math import logging from pathlib import Path import time import numpy as np from transformers import AutoTokenizer, AutoConfig import flexgen.apps.data_wrangle.utils.data_utils as data_utils import flexgen.apps.data_wrangle.utils.prompt_utils as prompt_utils from flexgen.apps.data_wrangle.utils import constants from flexgen.apps.data_wrangle.utils.utils import compute_metrics, setup_logger from flexgen.flex_opt import (Policy, OptLM, ExecutionEnv, CompressionConfig, str2bool) def add_flexgen_args(parser): parser.add_argument("--pad-to-seq-len", type=int) parser.add_argument("--model", type=str, default="facebook/opt-1.3b", help="The model name.") parser.add_argument("--path", type=str, default="~/opt_weights", help="The path to the model weights. If there are no cached weights, " "FlexGen will automatically download them from HuggingFace.") parser.add_argument("--run-path", type=str, default="runs") parser.add_argument("--offload-dir", type=str, default="~/flexgen_offload_dir", help="The directory to offload tensors. ") parser.add_argument("--gpu-batch-size", type=int, default=16) parser.add_argument("--num-gpu-batches", type=int, default=1) parser.add_argument("--percent", nargs="+", type=int, default=[100, 0, 100, 0, 100, 0], help="Six numbers. They are " "the percentage of weight on GPU, " "the percentage of weight on CPU, " "the percentage of attention cache on GPU, " "the percentage of attention cache on CPU, " "the percentage of activations on GPU, " "the percentage of activations on CPU") parser.add_argument("--pin-weight", type=str2bool, nargs="?", const=True, default=True) parser.add_argument("--cpu-cache-compute", action="store_true") parser.add_argument("--compress-weight", action="store_true", help="Whether to compress weight.") parser.add_argument("--compress-cache", action="store_true", help="Whether to compress cache.") The provided code snippet includes necessary dependencies for implementing the `parse_args` function. Write a Python function `def parse_args() -> argparse.Namespace` to solve the following problem: Generate args. Here is the function: def parse_args() -> argparse.Namespace: """Generate args.""" parser = argparse.ArgumentParser(description="Simple calculator") parser.add_argument( "--data_dir", type=str, help="Which data directory to run.", required=True, ) parser.add_argument( "--output_dir", type=str, help="Output directory.", default="outputs" ) parser.add_argument( "--cache_name", type=str, help="Manifest cache type.", default="sqlite", choices=["redis", "sqlite", "noop"], ) parser.add_argument( "--cache_connection", type=str, help="Manifest cache connection string.", default="fm_data_tasks.sqlite", ) parser.add_argument( "--client_name", type=str, help="Manifest client type.", default="openai", choices=["openai", "opt", "huggingface"], ) parser.add_argument( "--client_connection", type=str, help="Manifest client connection string.", default=None, ) parser.add_argument( "--run_tag", type=str, help="Tag for run saving.", default="default", ) parser.add_argument( "--overwrite_cache", action="store_true", help="Overwrite sqlite cache of input/output results.", ) parser.add_argument("--k", type=int, help="Number examples in prompt", default=1) parser.add_argument( "--sample_method", type=str, help="Example generation method", default="random", choices=["random", "manual", "validation_clusters"], ) parser.add_argument("--seed", type=int, default=1234) parser.add_argument( "--class_balanced", help="Class balance training data. Good for classification tasks \ with random prompts.", action="store_true", ) parser.add_argument( "--sep_tok", type=str, help="Separate for attr: val pairs in row. Default is '.'.", default=".", ) parser.add_argument( "--nan_tok", type=str, help="Token to represent nan entries. Default is 'nan'.", default="nan", ) parser.add_argument( "--num_run", type=int, help="Number examples to run through model.", default=-1, ) parser.add_argument( "--num_trials", type=int, help="Number trials to run. Results will be averaged with variance reported.", default=1, ) parser.add_argument( "--num_print", type=int, help="Number example prompts to print.", default=10, ) parser.add_argument( "--add_task_instruction", help="Add task instruction to the prompt before examples.", action="store_true", ) parser.add_argument("--task_instruction_idx", type=int, default=0) parser.add_argument("--do_test", help="Run on test file.", action="store_true") parser.add_argument( "--dry_run", help="Dry run. Do not actually ping model.", action="store_true" ) parser.add_argument( "--stop_token", help="Token to stop on for a given generated response", default="\n" ) # Model args parser.add_argument("--temperature", type=float, help="Temperature.", default=0.0) parser.add_argument( "--max_tokens", type=int, help="Max tokens to generate.", default=3 ) parser.add_argument( "--batch_run", help="Use FlexGen batch inference.", action="store_true" ) add_flexgen_args(parser) args = parser.parse_args() return args
Generate args.
10,173
import argparse from tqdm import tqdm import json import math import logging from pathlib import Path import time import numpy as np from transformers import AutoTokenizer, AutoConfig import flexgen.apps.data_wrangle.utils.data_utils as data_utils import flexgen.apps.data_wrangle.utils.prompt_utils as prompt_utils from flexgen.apps.data_wrangle.utils import constants from flexgen.apps.data_wrangle.utils.utils import compute_metrics, setup_logger from flexgen.flex_opt import (Policy, OptLM, ExecutionEnv, CompressionConfig, str2bool) logger = logging.getLogger(__name__) def get_tokenizer(name): if name == 'facebook/opt-175b': tokenizer = AutoTokenizer.from_pretrained('facebook/opt-30b', padding_side="left") else: tokenizer = AutoTokenizer.from_pretrained(name, padding_side="left") tokenizer.add_bos_token = False if 'galactica' in name: config = AutoConfig.from_pretrained(name) tokenizer.pad_token = config.pad_token_id tokenizer.eos_token = config.eos_token_id return tokenizer def compute_metrics(preds: List, golds: List, task: str): """Compute metrics.""" mets = {"tp": 0, "tn": 0, "fp": 0, "fn": 0, "crc": 0, "total": 0} for pred, label in zip(preds, golds): label = label.strip().lower() pred = pred.strip().lower() mets["total"] += 1 if task in { "data_imputation", "entity_matching", }: crc = pred == label elif task in {"entity_matching", "schema_matching", "error_detection_spelling"}: crc = pred.startswith(label) elif task in {"error_detection"}: pred = pred.split("\n\n")[-1] breakpoint() crc = pred.endswith(label) else: raise ValueError(f"Unknown task: {task}") # Measure equal accuracy for generation if crc: mets["crc"] += 1 if label == "yes": if crc: mets["tp"] += 1 else: mets["fn"] += 1 elif label == "no": if crc: mets["tn"] += 1 else: mets["fp"] += 1 prec = mets["tp"] / max(1, (mets["tp"] + mets["fp"])) rec = mets["tp"] / max(1, (mets["tp"] + mets["fn"])) acc = mets["crc"] / mets["total"] f1 = 2 * prec * rec / max(1, (prec + rec)) return prec, rec, acc, f1 class Policy: gpu_batch_size: int num_gpu_batches: int # percent = a means a% w_gpu_percent: float w_cpu_percent: float cache_gpu_percent: float cache_cpu_percent: float act_gpu_percent: float act_cpu_percent: float # Whether to overlap the I/O and compute overlap: bool # Whether to separate attention and mlp as two layers sep_layer: bool # Whether to use pinned memory for weights on CPU pin_weight: bool # Whether to compute attention on CPU cpu_cache_compute: bool # Sparsity of attention weights attn_sparsity: float # Compress weights with group-wise quantization compress_weight: bool comp_weight_config: CompressionConfig # Compress KV cache with group-wise quantization compress_cache: bool comp_cache_config: CompressionConfig def w_disk_percent(self): return 100 - self.w_gpu_percent - self.w_cpu_percent def cache_disk_percent(self): return 100 - self.cache_gpu_percent - self.cache_cpu_percent def act_disk_percent(self): return 100 - self.act_gpu_percent - self.act_cpu_percent class OptLM: def __init__(self, config: Union[str, OptConfig], env: ExecutionEnv, path: str, policy: Policy): if isinstance(config, str): config = get_opt_config(config) self.config = config self.env = env self.path = path self.policy = policy self.num_gpu_batches = policy.num_gpu_batches layers = [] layers.append(InputEmbed(self.config, self.env, self.policy)) for i in range(self.config.num_hidden_layers): if policy.sep_layer: layers.append(SelfAttention(self.config, self.env, self.policy, i)) layers.append(MLP(self.config, self.env, self.policy, i)) else: layers.append(TransformerLayer(self.config, self.env, self.policy, i)) layers.append(OutputEmbed(self.config, self.env, self.policy)) self.layers = layers self.num_layers = len(layers) if self.policy.act_gpu_percent == 100: self.act_home = self.env.gpu elif self.policy.act_cpu_percent == 100: self.act_home = self.env.cpu elif self.policy.act_disk_percent == 100: self.act_home = self.env.disk else: raise NotImplementedError() # CUDA streams self.load_weight_stream = torch.cuda.Stream() self.load_cache_stream = torch.cuda.Stream() self.store_cache_stream = torch.cuda.Stream() # Intermediate tensors # The following buffers store values used # for the i-th token, j-th layer, k-th gpu batch. num_layers, num_gpu_batches = self.num_layers, self.policy.num_gpu_batches # cache[j][k] self.cache_home = array_2d(num_layers, num_gpu_batches, ValueHolder) self.cache_read_buf = array_2d(num_layers, num_gpu_batches, ValueHolder) self.cache_write_buf = array_2d(num_layers, num_gpu_batches, ValueHolder) # weight[j] self.weight_read_buf = array_1d(num_layers, ValueHolder) # attention_mask[k] self.attention_mask = array_1d(num_gpu_batches, ValueHolder) self.task = None self.init_all_weights() def set_task(self, task): self.task = task for l in self.layers: l.set_task(task) def init_weight(self, j): expanded_path = os.path.abspath(os.path.expanduser( os.path.join(self.path, f"{self.config.name}-np"))) check_path = os.path.join(expanded_path, "decoder.embed_positions.weight") if not os.path.exists(check_path) and DUMMY_WEIGHT not in check_path: download_opt_weights(self.config.name, self.path) self.layers[j].init_weight(self.weight_home[j], expanded_path) def load_weight(self, i, j, k, overlap=True): # Handle corner cases if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load from weight_home to weight_read_buf if overlap: with torch.cuda.stream(self.load_weight_stream): self.layers[j].load_weight(self.weight_home[j], self.weight_read_buf[j], k) else: self.layers[j].load_weight(self.weight_home[j], self.weight_read_buf[j], k) def delete_weight(self, j, k): if k == 0: for x in self.weight_home[j].pop(): if isinstance(x, ValueHolder): for y in x.pop(): y.delete() else: x.delete() def init_cache(self, j, k): self.layers[j].init_cache_one_gpu_batch(self.cache_home[j][k]) def load_cache(self, i, j, k, overlap=True): # Handle corner cases if i == 0: # prefill, no cache return if k == self.num_gpu_batches: k = 0 j += 1 if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load from cache_home to cache_read_buf if overlap: with torch.cuda.stream(self.load_cache_stream): self.layers[j].load_cache(self.cache_home[j][k], self.cache_read_buf[j][k], i) else: self.layers[j].load_cache(self.cache_home[j][k], self.cache_read_buf[j][k], i) def store_cache(self, i, j, k, overlap=True): # Handle corner cases if k == -1: k = self.num_gpu_batches - 1 j -= 1 if j == -1: j = self.num_layers - 1 i -= 1 if i == -1: return if i == self.task.gen_len - 1: # last token, no need to store cache self.cache_write_buf[j][k].pop() return # Store cache_write_buf to cache_home # Delete cache_write_buf if overlap: with torch.cuda.stream(self.store_cache_stream): self.layers[j].store_cache(self.cache_home[j][k], self.cache_write_buf[j][k], i) else: self.layers[j].store_cache(self.cache_home[j][k], self.cache_write_buf[j][k], i) def delete_cache(self, j, k): v = self.cache_home[j][k].pop() if v: for x in v: x.delete() def load_hidden(self, i, j, k): # Handle corner cases if k == self.num_gpu_batches: k = 0 j += 1 if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load to hidden states buffers dst = self.layers[j].compute if j == 0: gpu_batch_size = self.policy.gpu_batch_size left, right = k * gpu_batch_size, (k + 1) * gpu_batch_size if i == 0: # load from the input ids val = dst.allocate((gpu_batch_size, self.task.prompt_len), np.int32) val.load_from_np(self.output_ids[left:right, :self.task.prompt_len]) else: # load from the last generated token pos = self.task.prompt_len + i val = dst.allocate((gpu_batch_size, 1), np.int32) val.load_from_np(self.output_ids[left:right, pos-1:pos]) else: # load from the last layer val = self.hidden[i][j-1][k].pop().move(dst) self.hidden[i][j][k].store(val) def store_hidden(self, i, j, k): # Handle corner cases if k == -1: k = self.num_gpu_batches - 1 j -= 1 if j == -1: j = self.num_layers - 1 i -= 1 if i == -1: return # Store to hidden states buffers if j == self.num_layers - 1: # store to output gpu_batch_size = self.policy.gpu_batch_size left, right = k * gpu_batch_size, (k + 1) * gpu_batch_size ids = self.hidden[i][j][k].pop().data.detach().cpu().numpy() pos = self.task.prompt_len + i if self.task.stop: stopped = self.stopped[left:right] self.output_ids[left:right, pos:pos+1] = np.where( stopped, self.config.pad_token_id, ids) stopped[:] = np.logical_or(stopped, ids == self.task.stop) else: self.output_ids[left:right, pos:pos+1] = ids else: # move to home x = self.hidden[i][j][k] if x.val: # x may already be moved due to overlapping x.val = x.val.move(self.act_home) def compute_layer(self, i, j, k): # Update the hidden in place # Clear the weight_read_buf if it is the last gpu batch # Clear the cache_read_buf # Run layer computation self.layers[j].forward(self.hidden[i][j][k], self.cache_read_buf[j][k], self.weight_read_buf[j], self.attention_mask[k], self.cache_write_buf[j][k], i, k) def sync(self): self.env.disk.synchronize() torch.cuda.synchronize() def init_all_weights(self): self.weight_home = array_1d(self.num_layers, ValueHolder) for j in range(self.num_layers): self.init_weight(j) def delete_all_weights(self): for j in range(self.num_layers): self.delete_weight(j, 0) def update_attention_mask(self, i, k): if i > 0: mask = self.attention_mask[k] assert mask.val is not None mask.val = mask.val.device.extend_attention_mask(mask.val, [True]) return gpu_batch_size = self.policy.gpu_batch_size left = k * gpu_batch_size right = left + gpu_batch_size input_ids = self.output_ids[left:right, :self.task.prompt_len] attention_compute = (self.env.cpu if self.policy.cpu_cache_compute else self.env.gpu) val = attention_compute.allocate( (self.policy.gpu_batch_size, self.task.prompt_len), bool) val.load_from_np((input_ids != self.config.pad_token_id)) self.attention_mask[k].store(val) def generate(self, inputs: Union[np.array, List[List[int]]], max_new_tokens: int = 32, do_sample: bool = False, temperature: float = 1.0, stop: Optional[int] = None, debug_mode: Optional[str] = None, cut_gen_len: Optional[int] = None, verbose: int = 0): task = Task( inputs=inputs, prompt_len=len(inputs[0]), gen_len=max_new_tokens, cut_gen_len=cut_gen_len, do_sample=do_sample, temperature=temperature, stop=stop, ) num_layers = self.num_layers num_gpu_batches = self.num_gpu_batches gpu_batch_size = self.policy.gpu_batch_size overlap = self.policy.overlap prompt_len, gen_len = task.prompt_len, task.gen_len self.execute_gen_len = task.cut_gen_len if task.cut_gen_len else task.gen_len # Output token ids self.output_ids = np.full((len(task.inputs), prompt_len + gen_len), self.config.pad_token_id, dtype=np.int32) self.stopped = np.zeros((len(task.inputs), 1), dtype=bool) self.output_ids[:, :prompt_len] = np.asarray(task.inputs) assert gpu_batch_size * num_gpu_batches == len(task.inputs) # Intermediate tensors # The following buffers store values used # for the i-th token, j-th layer, k-th gpu batch. num_layers, num_gpu_batches = self.num_layers, self.policy.num_gpu_batches for j in range(num_layers): for k in range(num_gpu_batches): self.cache_home[j][k].clear() self.cache_read_buf[j][k].clear() self.cache_write_buf[j][k].clear() for j in range(num_layers): self.weight_read_buf[j].clear() for k in range(num_gpu_batches): self.attention_mask[k].clear() self.hidden = array_3d(gen_len, num_layers, num_gpu_batches, ValueHolder) # Init cache self.set_task(task) for j in range(num_layers): for k in range(num_gpu_batches): self.init_cache(j, k) if self.policy.cpu_cache_compute: self.env.cpu.init_attention_compute_workspace(self.config, self.task, self.policy) # Generate if debug_mode is None: if not overlap: # No overlap, easy to understand, suitable for debugging self.generation_loop_normal() else: # Overlap I/O and compute if num_gpu_batches == 1: self.generation_loop_overlap_single_batch() else: self.generation_loop_overlap_multi_batch() elif debug_mode == "fewer_batch": # Run fewer layeres and batches for debugging if num_gpu_batches == 1: self.generation_loop_debug_single_batch() else: self.generation_loop_debug_multi_batch() elif debug_mode == "breakdown": # No overlap, fewer batches, execution time breakdown self.generation_loop_debug_normal() else: raise ValueError("Invalid debug mode: {debug_mode}") # Delete cache for j in range(num_layers): for k in range(num_gpu_batches): self.delete_cache(j, k) if self.policy.cpu_cache_compute: self.env.cpu.del_attention_compute_workspace() return self.output_ids def generation_loop_normal(self): for i in range(self.execute_gen_len): timers("generate").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): for k in range(self.num_gpu_batches): self.load_weight(i, j, k, overlap=False) for k in range(self.num_gpu_batches): self.load_cache(i, j, k, overlap=False) self.load_hidden(i, j, k) self.compute_layer(i, j, k) self.store_hidden(i, j, k) self.store_cache(i, j, k, overlap=False) timers("generate").stop() def generation_loop_debug_normal(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill_total").reset() timers("decoding_gpu_batch").reset() timers("load_weight").reset() timers("load_cache_prefill").reset() timers("load_cache_decoding").reset() timers("store_cache_prefill").reset() timers("store_cache_decoding").reset() timers("compute_layer_prefill").reset() timers("compute_layer_decoding").reset() load_weight_timer = timers("load_weight") for i in range(self.execute_gen_len): if i == 0: timers("prefill_total").start() load_cache_timer = timers("load_cache_prefill") store_cache_timer = timers("store_cache_prefill") compute_layer_timer = timers("compute_layer_prefill") else: load_cache_timer = timers("load_cache_decoding") store_cache_timer = timers("store_cache_decoding") compute_layer_timer = timers("compute_layer_decoding") for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() load_weight_timer.start(self.sync) for k in range(self.num_gpu_batches): self.load_weight(i, j, k) load_weight_timer.stop(self.sync) for k in range(self.num_gpu_batches): load_cache_timer.start(self.sync) self.load_cache(i, j, k) load_cache_timer.stop(self.sync) self.load_hidden(i, j, k) compute_layer_timer.start(self.sync) self.compute_layer(i, j, k) compute_layer_timer.stop(self.sync) self.store_hidden(i, j, k) store_cache_timer.start(self.sync) self.store_cache(i, j, k) store_cache_timer.stop(self.sync) if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill_total").stop(self.sync) # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill_total").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) # Debug the costs of individual functions print(f"#layers: {self.num_layers}") print(f"#batches prefill: " f"{self.num_layers * self.num_gpu_batches}") print(f"#batches decoding: " f"{(self.task.gen_len - 1) * self.num_layers * self.num_gpu_batches}") print(f"load_weight (per-layer)" f": {np.mean(timers('load_weight').costs):.6f} s") for stage in ["prefill", "decoding"]: for func in ["load_cache", "store_cache", "compute_layer"]: name = func + "_" + stage costs = timers(name).costs print(f"{name:22s} (per-batch): {np.mean(costs):.6f} s") def generation_loop_overlap_single_batch(self): # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.sync() # Generate for i in range(self.execute_gen_len): timers("generate").start() self.update_attention_mask(i, 0) for j in range(self.num_layers): self.load_weight(i, j+1, 0) self.load_cache(i, j+1, 0) self.load_hidden(i, j, 0) self.compute_layer(i, j, 0) self.store_cache(i, j-1, 0) self.store_hidden(i, j, 0) self.sync() timers("generate").stop() if self.task.stop and np.all(self.stopped): break def generation_loop_overlap_multi_batch(self): # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.load_hidden(0, 0, 0) self.sync() # Generate for i in range(self.execute_gen_len): timers("generate").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): for k in range(self.num_gpu_batches): self.load_weight(i, j+1, k) self.load_cache(i, j, k+1) self.store_hidden(i, j, k-1) self.load_hidden(i, j, k+1) self.compute_layer(i, j, k) self.store_cache(i, j, k-1) self.sync() timers("generate").stop() # Epilogue self.store_hidden( self.execute_gen_len-1, self.num_layers-1, self.num_gpu_batches-1) def generation_loop_debug_single_batch(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill").reset() timers("decoding_gpu_batch").reset() # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.sync() # Generate for i in range(self.execute_gen_len): if i == 0: timers("prefill").start() self.update_attention_mask(i, 0) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() self.load_weight(i, j+1, 0) self.load_cache(i, j+1, 0) self.load_hidden(i, j, 0) self.compute_layer(i, j, 0) self.store_cache(i, j-1, 0) self.store_hidden(i, j, 0) self.sync() if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill").stop() # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) def generation_loop_debug_multi_batch(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill").reset() timers("decoding_gpu_batch").reset() # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.load_hidden(0, 0, 0) self.sync() # Generate for i in range(self.execute_gen_len): if i == 0: timers("prefill").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() for k in range(self.num_gpu_batches): self.load_weight(i, j+1, k) self.load_cache(i, j, k+1) self.store_hidden(i, j, k-1) self.load_hidden(i, j, k+1) self.compute_layer(i, j, k) self.store_cache(i, j, k-1) self.sync() if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill").stop() # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) def __del__(self): self.delete_all_weights() def single_query_test(args, task_instruction, test_data, task, pd_data_files, test_file): # Initialize environment tokenizer = get_tokenizer(args.model) env = ExecutionEnv.create(args.offload_dir) # Offloading policy policy = Policy(1, 1, args.percent[0], args.percent[1], args.percent[2], args.percent[3], args.percent[4], args.percent[5], overlap=True, sep_layer=True, pin_weight=args.pin_weight, cpu_cache_compute=args.cpu_cache_compute, attn_sparsity=1.0, compress_weight=args.compress_weight, comp_weight_config=CompressionConfig( num_bits=4, group_size=64, group_dim=0, symmetric=False), compress_cache=args.compress_cache, comp_cache_config=CompressionConfig( num_bits=4, group_size=64, group_dim=2, symmetric=False)) logger.info(f"Init weights begin.") tic = time.time() model = OptLM(args.model, env, args.path, policy) logger.info(f"Init weights end. Elapsed: {time.time() - tic:.2f} s") if args.add_task_instruction: prompt = lambda x: f"{task_instruction} {x}" else: prompt = lambda x: f"{x}" trial_metrics = {"prec": [], "rec": [], "f1": [], "acc": []} saved_prefix = None tic = time.time() for trial_num in range(args.num_trials): np.random.seed(args.seed + trial_num) queries = [] for _, row in test_data.iterrows(): serialized_r = row["text"] if args.sample_method == "manual": prefix_exs = prompt_utils.get_manual_prompt(args.data_dir, row) elif args.sample_method == "validation_clusters": if saved_prefix is None: logger.info("Generating validation cluster prompt.") saved_prefix = prompt_utils.get_validation_prompt( args.validation_path, num_examples=args.k, task=task, ) prefix_exs = saved_prefix else: if saved_prefix is None: saved_prefix = prompt_utils.get_random_prompt( pd_data_files["train"], num_examples=args.k ) prefix_exs = saved_prefix queries.append((prefix_exs + "\n" + serialized_r).strip()) gt = test_data["label_str"] preds = [] idx = 0 for _ in range(args.num_print): logger.info(prompt(queries[idx])) tic = time.time() input_ids_tmp = tokenizer(prompt(queries[idx]), padding="max_length", return_tensors="np", max_length=args.pad_to_seq_len).input_ids logger.info(input_ids_tmp.shape) output_ids_tmp = model.generate(input_ids_tmp, do_sample=True, temperature=args.temperature, max_new_tokens=args.max_tokens, stop=args.stop_token) input_strs = tokenizer.batch_decode(input_ids_tmp, skip_special_tokens=True) output_strs = tokenizer.batch_decode(output_ids_tmp, skip_special_tokens=True) anwsers = [ output_strs[i][len(input_strs[i]):] for i in range(len(input_strs))] logger.info(f"====> {anwsers[0]} <====") preds.extend(anwsers) idx += 1 logger.info(f"Current Inference query elapsed: {time.time() - tic:.2f} s") # Save trial predictions save_data = test_data.iloc[:args.num_print].copy(deep=True).reset_index() gt = gt[:args.num_print] save_data["preds"] = preds save_data["queries"] = queries[:args.num_print] prec, rec, acc, f1 = compute_metrics(preds, gt, task) logger.info( f"Metrics Trial {trial_num}\n" f"Prec: {prec:.3f} Recall: {rec:.3f} Acc: {acc:.3f} F1: {f1:.3f}" ) trial_metrics["rec"].append(rec) trial_metrics["prec"].append(prec) trial_metrics["acc"].append(acc) trial_metrics["f1"].append(f1) output_file = ( Path(args.output_dir) / f"{Path(args.data_dir).stem}" / f"{test_file}" / f"{args.run_tag}" / f"{args.k}k" f"_{int(args.add_task_instruction)}inst" f"_{int(args.class_balanced)}cb" f"_{args.sample_method}" f"_{args.model}" f"_{args.num_print}run" f"_{int(args.dry_run)}dry" / f"trial_{trial_num}.feather" ) output_file.parent.mkdir(parents=True, exist_ok=True) logger.info(f"Saved to {output_file}") save_data.to_feather(output_file) for k, values in list(trial_metrics.items()): trial_metrics[f"{k}_avg"] = np.average(values) trial_metrics[f"{k}_std"] = np.std(values) output_metrics = output_file.parent / "metrics.json" json.dump(trial_metrics, open(output_metrics, "w")) logger.info(f"Final Metrics {json.dumps(trial_metrics, indent=4)}") logger.info(f"Metrics dumped to {output_metrics}") # Shutdown logger.info("Shutdown FlexGen...") env.close_copy_threads()
null
10,174
import argparse from tqdm import tqdm import json import math import logging from pathlib import Path import time import numpy as np from transformers import AutoTokenizer, AutoConfig import flexgen.apps.data_wrangle.utils.data_utils as data_utils import flexgen.apps.data_wrangle.utils.prompt_utils as prompt_utils from flexgen.apps.data_wrangle.utils import constants from flexgen.apps.data_wrangle.utils.utils import compute_metrics, setup_logger from flexgen.flex_opt import (Policy, OptLM, ExecutionEnv, CompressionConfig, str2bool) logger = logging.getLogger(__name__) def get_tokenizer(name): if name == 'facebook/opt-175b': tokenizer = AutoTokenizer.from_pretrained('facebook/opt-30b', padding_side="left") else: tokenizer = AutoTokenizer.from_pretrained(name, padding_side="left") tokenizer.add_bos_token = False if 'galactica' in name: config = AutoConfig.from_pretrained(name) tokenizer.pad_token = config.pad_token_id tokenizer.eos_token = config.eos_token_id return tokenizer def compute_metrics(preds: List, golds: List, task: str): """Compute metrics.""" mets = {"tp": 0, "tn": 0, "fp": 0, "fn": 0, "crc": 0, "total": 0} for pred, label in zip(preds, golds): label = label.strip().lower() pred = pred.strip().lower() mets["total"] += 1 if task in { "data_imputation", "entity_matching", }: crc = pred == label elif task in {"entity_matching", "schema_matching", "error_detection_spelling"}: crc = pred.startswith(label) elif task in {"error_detection"}: pred = pred.split("\n\n")[-1] breakpoint() crc = pred.endswith(label) else: raise ValueError(f"Unknown task: {task}") # Measure equal accuracy for generation if crc: mets["crc"] += 1 if label == "yes": if crc: mets["tp"] += 1 else: mets["fn"] += 1 elif label == "no": if crc: mets["tn"] += 1 else: mets["fp"] += 1 prec = mets["tp"] / max(1, (mets["tp"] + mets["fp"])) rec = mets["tp"] / max(1, (mets["tp"] + mets["fn"])) acc = mets["crc"] / mets["total"] f1 = 2 * prec * rec / max(1, (prec + rec)) return prec, rec, acc, f1 class Policy: gpu_batch_size: int num_gpu_batches: int # percent = a means a% w_gpu_percent: float w_cpu_percent: float cache_gpu_percent: float cache_cpu_percent: float act_gpu_percent: float act_cpu_percent: float # Whether to overlap the I/O and compute overlap: bool # Whether to separate attention and mlp as two layers sep_layer: bool # Whether to use pinned memory for weights on CPU pin_weight: bool # Whether to compute attention on CPU cpu_cache_compute: bool # Sparsity of attention weights attn_sparsity: float # Compress weights with group-wise quantization compress_weight: bool comp_weight_config: CompressionConfig # Compress KV cache with group-wise quantization compress_cache: bool comp_cache_config: CompressionConfig def w_disk_percent(self): return 100 - self.w_gpu_percent - self.w_cpu_percent def cache_disk_percent(self): return 100 - self.cache_gpu_percent - self.cache_cpu_percent def act_disk_percent(self): return 100 - self.act_gpu_percent - self.act_cpu_percent class OptLM: def __init__(self, config: Union[str, OptConfig], env: ExecutionEnv, path: str, policy: Policy): if isinstance(config, str): config = get_opt_config(config) self.config = config self.env = env self.path = path self.policy = policy self.num_gpu_batches = policy.num_gpu_batches layers = [] layers.append(InputEmbed(self.config, self.env, self.policy)) for i in range(self.config.num_hidden_layers): if policy.sep_layer: layers.append(SelfAttention(self.config, self.env, self.policy, i)) layers.append(MLP(self.config, self.env, self.policy, i)) else: layers.append(TransformerLayer(self.config, self.env, self.policy, i)) layers.append(OutputEmbed(self.config, self.env, self.policy)) self.layers = layers self.num_layers = len(layers) if self.policy.act_gpu_percent == 100: self.act_home = self.env.gpu elif self.policy.act_cpu_percent == 100: self.act_home = self.env.cpu elif self.policy.act_disk_percent == 100: self.act_home = self.env.disk else: raise NotImplementedError() # CUDA streams self.load_weight_stream = torch.cuda.Stream() self.load_cache_stream = torch.cuda.Stream() self.store_cache_stream = torch.cuda.Stream() # Intermediate tensors # The following buffers store values used # for the i-th token, j-th layer, k-th gpu batch. num_layers, num_gpu_batches = self.num_layers, self.policy.num_gpu_batches # cache[j][k] self.cache_home = array_2d(num_layers, num_gpu_batches, ValueHolder) self.cache_read_buf = array_2d(num_layers, num_gpu_batches, ValueHolder) self.cache_write_buf = array_2d(num_layers, num_gpu_batches, ValueHolder) # weight[j] self.weight_read_buf = array_1d(num_layers, ValueHolder) # attention_mask[k] self.attention_mask = array_1d(num_gpu_batches, ValueHolder) self.task = None self.init_all_weights() def set_task(self, task): self.task = task for l in self.layers: l.set_task(task) def init_weight(self, j): expanded_path = os.path.abspath(os.path.expanduser( os.path.join(self.path, f"{self.config.name}-np"))) check_path = os.path.join(expanded_path, "decoder.embed_positions.weight") if not os.path.exists(check_path) and DUMMY_WEIGHT not in check_path: download_opt_weights(self.config.name, self.path) self.layers[j].init_weight(self.weight_home[j], expanded_path) def load_weight(self, i, j, k, overlap=True): # Handle corner cases if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load from weight_home to weight_read_buf if overlap: with torch.cuda.stream(self.load_weight_stream): self.layers[j].load_weight(self.weight_home[j], self.weight_read_buf[j], k) else: self.layers[j].load_weight(self.weight_home[j], self.weight_read_buf[j], k) def delete_weight(self, j, k): if k == 0: for x in self.weight_home[j].pop(): if isinstance(x, ValueHolder): for y in x.pop(): y.delete() else: x.delete() def init_cache(self, j, k): self.layers[j].init_cache_one_gpu_batch(self.cache_home[j][k]) def load_cache(self, i, j, k, overlap=True): # Handle corner cases if i == 0: # prefill, no cache return if k == self.num_gpu_batches: k = 0 j += 1 if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load from cache_home to cache_read_buf if overlap: with torch.cuda.stream(self.load_cache_stream): self.layers[j].load_cache(self.cache_home[j][k], self.cache_read_buf[j][k], i) else: self.layers[j].load_cache(self.cache_home[j][k], self.cache_read_buf[j][k], i) def store_cache(self, i, j, k, overlap=True): # Handle corner cases if k == -1: k = self.num_gpu_batches - 1 j -= 1 if j == -1: j = self.num_layers - 1 i -= 1 if i == -1: return if i == self.task.gen_len - 1: # last token, no need to store cache self.cache_write_buf[j][k].pop() return # Store cache_write_buf to cache_home # Delete cache_write_buf if overlap: with torch.cuda.stream(self.store_cache_stream): self.layers[j].store_cache(self.cache_home[j][k], self.cache_write_buf[j][k], i) else: self.layers[j].store_cache(self.cache_home[j][k], self.cache_write_buf[j][k], i) def delete_cache(self, j, k): v = self.cache_home[j][k].pop() if v: for x in v: x.delete() def load_hidden(self, i, j, k): # Handle corner cases if k == self.num_gpu_batches: k = 0 j += 1 if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load to hidden states buffers dst = self.layers[j].compute if j == 0: gpu_batch_size = self.policy.gpu_batch_size left, right = k * gpu_batch_size, (k + 1) * gpu_batch_size if i == 0: # load from the input ids val = dst.allocate((gpu_batch_size, self.task.prompt_len), np.int32) val.load_from_np(self.output_ids[left:right, :self.task.prompt_len]) else: # load from the last generated token pos = self.task.prompt_len + i val = dst.allocate((gpu_batch_size, 1), np.int32) val.load_from_np(self.output_ids[left:right, pos-1:pos]) else: # load from the last layer val = self.hidden[i][j-1][k].pop().move(dst) self.hidden[i][j][k].store(val) def store_hidden(self, i, j, k): # Handle corner cases if k == -1: k = self.num_gpu_batches - 1 j -= 1 if j == -1: j = self.num_layers - 1 i -= 1 if i == -1: return # Store to hidden states buffers if j == self.num_layers - 1: # store to output gpu_batch_size = self.policy.gpu_batch_size left, right = k * gpu_batch_size, (k + 1) * gpu_batch_size ids = self.hidden[i][j][k].pop().data.detach().cpu().numpy() pos = self.task.prompt_len + i if self.task.stop: stopped = self.stopped[left:right] self.output_ids[left:right, pos:pos+1] = np.where( stopped, self.config.pad_token_id, ids) stopped[:] = np.logical_or(stopped, ids == self.task.stop) else: self.output_ids[left:right, pos:pos+1] = ids else: # move to home x = self.hidden[i][j][k] if x.val: # x may already be moved due to overlapping x.val = x.val.move(self.act_home) def compute_layer(self, i, j, k): # Update the hidden in place # Clear the weight_read_buf if it is the last gpu batch # Clear the cache_read_buf # Run layer computation self.layers[j].forward(self.hidden[i][j][k], self.cache_read_buf[j][k], self.weight_read_buf[j], self.attention_mask[k], self.cache_write_buf[j][k], i, k) def sync(self): self.env.disk.synchronize() torch.cuda.synchronize() def init_all_weights(self): self.weight_home = array_1d(self.num_layers, ValueHolder) for j in range(self.num_layers): self.init_weight(j) def delete_all_weights(self): for j in range(self.num_layers): self.delete_weight(j, 0) def update_attention_mask(self, i, k): if i > 0: mask = self.attention_mask[k] assert mask.val is not None mask.val = mask.val.device.extend_attention_mask(mask.val, [True]) return gpu_batch_size = self.policy.gpu_batch_size left = k * gpu_batch_size right = left + gpu_batch_size input_ids = self.output_ids[left:right, :self.task.prompt_len] attention_compute = (self.env.cpu if self.policy.cpu_cache_compute else self.env.gpu) val = attention_compute.allocate( (self.policy.gpu_batch_size, self.task.prompt_len), bool) val.load_from_np((input_ids != self.config.pad_token_id)) self.attention_mask[k].store(val) def generate(self, inputs: Union[np.array, List[List[int]]], max_new_tokens: int = 32, do_sample: bool = False, temperature: float = 1.0, stop: Optional[int] = None, debug_mode: Optional[str] = None, cut_gen_len: Optional[int] = None, verbose: int = 0): task = Task( inputs=inputs, prompt_len=len(inputs[0]), gen_len=max_new_tokens, cut_gen_len=cut_gen_len, do_sample=do_sample, temperature=temperature, stop=stop, ) num_layers = self.num_layers num_gpu_batches = self.num_gpu_batches gpu_batch_size = self.policy.gpu_batch_size overlap = self.policy.overlap prompt_len, gen_len = task.prompt_len, task.gen_len self.execute_gen_len = task.cut_gen_len if task.cut_gen_len else task.gen_len # Output token ids self.output_ids = np.full((len(task.inputs), prompt_len + gen_len), self.config.pad_token_id, dtype=np.int32) self.stopped = np.zeros((len(task.inputs), 1), dtype=bool) self.output_ids[:, :prompt_len] = np.asarray(task.inputs) assert gpu_batch_size * num_gpu_batches == len(task.inputs) # Intermediate tensors # The following buffers store values used # for the i-th token, j-th layer, k-th gpu batch. num_layers, num_gpu_batches = self.num_layers, self.policy.num_gpu_batches for j in range(num_layers): for k in range(num_gpu_batches): self.cache_home[j][k].clear() self.cache_read_buf[j][k].clear() self.cache_write_buf[j][k].clear() for j in range(num_layers): self.weight_read_buf[j].clear() for k in range(num_gpu_batches): self.attention_mask[k].clear() self.hidden = array_3d(gen_len, num_layers, num_gpu_batches, ValueHolder) # Init cache self.set_task(task) for j in range(num_layers): for k in range(num_gpu_batches): self.init_cache(j, k) if self.policy.cpu_cache_compute: self.env.cpu.init_attention_compute_workspace(self.config, self.task, self.policy) # Generate if debug_mode is None: if not overlap: # No overlap, easy to understand, suitable for debugging self.generation_loop_normal() else: # Overlap I/O and compute if num_gpu_batches == 1: self.generation_loop_overlap_single_batch() else: self.generation_loop_overlap_multi_batch() elif debug_mode == "fewer_batch": # Run fewer layeres and batches for debugging if num_gpu_batches == 1: self.generation_loop_debug_single_batch() else: self.generation_loop_debug_multi_batch() elif debug_mode == "breakdown": # No overlap, fewer batches, execution time breakdown self.generation_loop_debug_normal() else: raise ValueError("Invalid debug mode: {debug_mode}") # Delete cache for j in range(num_layers): for k in range(num_gpu_batches): self.delete_cache(j, k) if self.policy.cpu_cache_compute: self.env.cpu.del_attention_compute_workspace() return self.output_ids def generation_loop_normal(self): for i in range(self.execute_gen_len): timers("generate").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): for k in range(self.num_gpu_batches): self.load_weight(i, j, k, overlap=False) for k in range(self.num_gpu_batches): self.load_cache(i, j, k, overlap=False) self.load_hidden(i, j, k) self.compute_layer(i, j, k) self.store_hidden(i, j, k) self.store_cache(i, j, k, overlap=False) timers("generate").stop() def generation_loop_debug_normal(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill_total").reset() timers("decoding_gpu_batch").reset() timers("load_weight").reset() timers("load_cache_prefill").reset() timers("load_cache_decoding").reset() timers("store_cache_prefill").reset() timers("store_cache_decoding").reset() timers("compute_layer_prefill").reset() timers("compute_layer_decoding").reset() load_weight_timer = timers("load_weight") for i in range(self.execute_gen_len): if i == 0: timers("prefill_total").start() load_cache_timer = timers("load_cache_prefill") store_cache_timer = timers("store_cache_prefill") compute_layer_timer = timers("compute_layer_prefill") else: load_cache_timer = timers("load_cache_decoding") store_cache_timer = timers("store_cache_decoding") compute_layer_timer = timers("compute_layer_decoding") for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() load_weight_timer.start(self.sync) for k in range(self.num_gpu_batches): self.load_weight(i, j, k) load_weight_timer.stop(self.sync) for k in range(self.num_gpu_batches): load_cache_timer.start(self.sync) self.load_cache(i, j, k) load_cache_timer.stop(self.sync) self.load_hidden(i, j, k) compute_layer_timer.start(self.sync) self.compute_layer(i, j, k) compute_layer_timer.stop(self.sync) self.store_hidden(i, j, k) store_cache_timer.start(self.sync) self.store_cache(i, j, k) store_cache_timer.stop(self.sync) if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill_total").stop(self.sync) # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill_total").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) # Debug the costs of individual functions print(f"#layers: {self.num_layers}") print(f"#batches prefill: " f"{self.num_layers * self.num_gpu_batches}") print(f"#batches decoding: " f"{(self.task.gen_len - 1) * self.num_layers * self.num_gpu_batches}") print(f"load_weight (per-layer)" f": {np.mean(timers('load_weight').costs):.6f} s") for stage in ["prefill", "decoding"]: for func in ["load_cache", "store_cache", "compute_layer"]: name = func + "_" + stage costs = timers(name).costs print(f"{name:22s} (per-batch): {np.mean(costs):.6f} s") def generation_loop_overlap_single_batch(self): # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.sync() # Generate for i in range(self.execute_gen_len): timers("generate").start() self.update_attention_mask(i, 0) for j in range(self.num_layers): self.load_weight(i, j+1, 0) self.load_cache(i, j+1, 0) self.load_hidden(i, j, 0) self.compute_layer(i, j, 0) self.store_cache(i, j-1, 0) self.store_hidden(i, j, 0) self.sync() timers("generate").stop() if self.task.stop and np.all(self.stopped): break def generation_loop_overlap_multi_batch(self): # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.load_hidden(0, 0, 0) self.sync() # Generate for i in range(self.execute_gen_len): timers("generate").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): for k in range(self.num_gpu_batches): self.load_weight(i, j+1, k) self.load_cache(i, j, k+1) self.store_hidden(i, j, k-1) self.load_hidden(i, j, k+1) self.compute_layer(i, j, k) self.store_cache(i, j, k-1) self.sync() timers("generate").stop() # Epilogue self.store_hidden( self.execute_gen_len-1, self.num_layers-1, self.num_gpu_batches-1) def generation_loop_debug_single_batch(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill").reset() timers("decoding_gpu_batch").reset() # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.sync() # Generate for i in range(self.execute_gen_len): if i == 0: timers("prefill").start() self.update_attention_mask(i, 0) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() self.load_weight(i, j+1, 0) self.load_cache(i, j+1, 0) self.load_hidden(i, j, 0) self.compute_layer(i, j, 0) self.store_cache(i, j-1, 0) self.store_hidden(i, j, 0) self.sync() if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill").stop() # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) def generation_loop_debug_multi_batch(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill").reset() timers("decoding_gpu_batch").reset() # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.load_hidden(0, 0, 0) self.sync() # Generate for i in range(self.execute_gen_len): if i == 0: timers("prefill").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() for k in range(self.num_gpu_batches): self.load_weight(i, j+1, k) self.load_cache(i, j, k+1) self.store_hidden(i, j, k-1) self.load_hidden(i, j, k+1) self.compute_layer(i, j, k) self.store_cache(i, j, k-1) self.sync() if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill").stop() # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) def __del__(self): self.delete_all_weights() def batch_query_test(args, task_instruction, test_data, task, pd_data_files, test_file): # Initialize environment tokenizer = get_tokenizer(args.model) env = ExecutionEnv.create(args.offload_dir) # Offloading policy policy = Policy(args.gpu_batch_size, args.num_gpu_batches, args.percent[0], args.percent[1], args.percent[2], args.percent[3], args.percent[4], args.percent[5], overlap=True, sep_layer=True, pin_weight=args.pin_weight, cpu_cache_compute=args.cpu_cache_compute, attn_sparsity=1.0, compress_weight=args.compress_weight, comp_weight_config=CompressionConfig( num_bits=4, group_size=64, group_dim=0, symmetric=False), compress_cache=args.compress_cache, comp_cache_config=CompressionConfig( num_bits=4, group_size=64, group_dim=2, symmetric=False)) logger.info(f"Init weights begin.") tic = time.time() model = OptLM(args.model, env, args.path, policy) logger.info(f"Init weights end. Elapsed: {time.time() - tic:.2f} s.") if args.add_task_instruction: prompt = lambda x: f"{task_instruction} {x}" else: prompt = lambda x: f"{x}" trial_metrics = {"prec": [], "rec": [], "f1": [], "acc": [], "total_time": [], "output_throughput": [], "total_throughput": []} saved_prefix = None for trial_num in range(args.num_trials): np.random.seed(args.seed + trial_num) queries = [] for _, row in test_data.iterrows(): serialized_r = row["text"] if args.sample_method == "manual": prefix_exs = prompt_utils.get_manual_prompt(args.data_dir, row) elif args.sample_method == "validation_clusters": if saved_prefix is None: logger.info("Generating validation cluster prompt.") saved_prefix = prompt_utils.get_validation_prompt( args.validation_path, num_examples=args.k, task=task, ) prefix_exs = saved_prefix else: if saved_prefix is None: saved_prefix = prompt_utils.get_random_prompt( pd_data_files["train"], num_examples=args.k ) prefix_exs = saved_prefix queries.append((prefix_exs + "\n" + serialized_r).strip()) gt = test_data["label_str"] preds = [] idx = 0 max_prompt_seq_length = 0 prompt_strs = [] for _ in range(args.num_run): # if idx == 0: # logger.info(f"This is a sample prompt: {prompt(queries[idx])}") prompt_strs.append(prompt(queries[idx])) current_prompt_tmp = tokenizer(prompt(queries[idx]), padding="max_length", return_tensors="np", max_length=args.pad_to_seq_len).input_ids # logger.info(f"Current prompt <{idx}> length: {current_prompt_tmp.shape[1]}") max_prompt_seq_length = max(max_prompt_seq_length, current_prompt_tmp.shape[1]) idx += 1 logger.info(f"max_prompt_seq_length: {max_prompt_seq_length}") tic = time.time() input_ids = tokenizer(prompt_strs, padding="max_length", return_tensors="np", max_length=max_prompt_seq_length).input_ids output_ids = [] flexgen_batch_size = args.gpu_batch_size*args.num_gpu_batches num_batched_run = math.floor(args.num_run/flexgen_batch_size) args.num_run = num_batched_run * flexgen_batch_size input_ids = input_ids[0:args.num_run] for i in tqdm(range(num_batched_run)): input_ids_tmp = input_ids[i*flexgen_batch_size: (i+1)*flexgen_batch_size] output_ids_tmp = model.generate(input_ids_tmp, do_sample=True, temperature=args.temperature, max_new_tokens=args.max_tokens, stop=args.stop_token) output_ids.extend(output_ids_tmp) toc = time.time() input_strs = tokenizer.batch_decode(input_ids, skip_special_tokens=True) output_strs = tokenizer.batch_decode(output_ids, skip_special_tokens=True) preds = [output_strs[i][len(input_strs[i]):] for i in range(len(input_strs))] total_time = time.time() - tic total_prompt_tokens = args.num_run * max_prompt_seq_length total_generate_tokens = args.num_run * args.max_tokens output_throughput = total_generate_tokens/total_time total_throughput = (total_prompt_tokens+total_generate_tokens)/total_time logger.info(f"Batch inference run end. Elapsed: {total_time:.2f} s;") logger.info(f"Output throughput: {output_throughput:.2f} token/s;") logger.info(f"Total throughput: {total_throughput:.2f} token/s;") # Save trial predictions save_data = test_data.iloc[:args.num_run].copy(deep=True).reset_index() gt = gt[:args.num_run] save_data["preds"] = preds save_data["queries"] = queries[:args.num_run] prec, rec, acc, f1 = compute_metrics(preds, gt, task) logger.info( f"Metrics Trial {trial_num}\n" f"Prec: {prec:.3f} Recall: {rec:.3f} Acc: {acc:.3f} F1: {f1:.3f} \n" f"<FlexGen> time: {total_time:.3f} \n" f"<FlexGen> output throughput: {output_throughput:.3f} \n" f"<FlexGen> total throughput: {total_throughput:.3f}" ) trial_metrics["rec"].append(rec) trial_metrics["prec"].append(prec) trial_metrics["acc"].append(acc) trial_metrics["f1"].append(f1) trial_metrics["total_time"].append(total_time) trial_metrics["output_throughput"].append(output_throughput) trial_metrics["total_throughput"].append(total_throughput) output_file = ( Path(args.output_dir) / f"{Path(args.data_dir).stem}" / f"{test_file}" / f"{args.run_tag}" / f"{args.k}k" f"_{int(args.add_task_instruction)}inst" f"_{int(args.class_balanced)}cb" f"_{args.sample_method}" f"_{args.model}" f"_{args.num_run}run" f"_{int(args.dry_run)}dry" / f"trial_{trial_num}.feather" ) output_file.parent.mkdir(parents=True, exist_ok=True) logger.info(f"Saved to {output_file}") save_data.to_feather(output_file) for k, values in list(trial_metrics.items()): trial_metrics[f"{k}_avg"] = np.average(values) trial_metrics[f"{k}_std"] = np.std(values) output_metrics = output_file.parent / "metrics.json" json.dump(trial_metrics, open(output_metrics, "w")) logger.info(f"Final Metrics {json.dumps(trial_metrics, indent=4)}") logger.info(f"Metrics dumped to {output_metrics}") # Shutdown logger.info("Shutdown FlexGen...") env.close_copy_threads()
null
10,175
import logging from pathlib import Path from typing import List from rich.logging import RichHandler import logging The provided code snippet includes necessary dependencies for implementing the `setup_logger` function. Write a Python function `def setup_logger(log_dir: str)` to solve the following problem: Create log directory and logger. Here is the function: def setup_logger(log_dir: str): """Create log directory and logger.""" Path(log_dir).mkdir(exist_ok=True, parents=True) log_path = str(Path(log_dir) / "log.txt") handlers = [logging.FileHandler(log_path), RichHandler(rich_tracebacks=True)] logging.basicConfig( level=logging.INFO, format="%(asctime)s [%(module)s] [%(levelname)s] %(message)s", handlers=handlers, )
Create log directory and logger.
10,176
import logging from functools import partial from pathlib import Path from typing import Dict, List import pandas as pd from flexgen.apps.data_wrangle.utils import constants logger = logging.getLogger(__name__) def read_raw_data( data_dir: str, add_instruction: bool = False, task_instruction_idx: int = 0, sep_tok: str = ".", nan_tok: str = "nan", ): """Read in data where each directory is unique for a task.""" data_files_sep = {"test": {}, "train": {}, "validation": {}} logger.info(f"Processing {data_dir}") if data_dir not in constants.DATA2TASK: raise ValueError( f"{data_dir} not one of {constants.DATA2TASK.keys()}. Make sure to set DATASET_PATH." ) task = constants.DATA2TASK[data_dir] instruction = constants.DATA2INSTRUCT[data_dir] suffix = constants.DATA2SUFFIX[data_dir] cols_to_drop = constants.DATA2DROPCOLS[data_dir] col_renaming = constants.DATA2COLREMAP[data_dir] data_dir_p = Path(data_dir) if task == "entity_matching": train_file = data_dir_p / "train.csv" valid_file = data_dir_p / "valid.csv" test_file = data_dir_p / "test.csv" tableA_file = data_dir_p / "tableA.csv" tableB_file = data_dir_p / "tableB.csv" tableA = pd.read_csv(tableA_file) tableB = pd.read_csv(tableB_file) label_col = "label" read_data_func = partial( read_blocked_pairs, tableA=tableA, tableB=tableB, cols_to_drop=cols_to_drop, col_renaming=col_renaming, add_instruction=add_instruction, instruction=instruction, suffix=suffix, prod_name=constants.MATCH_PROD_NAME[data_dir], sep_tok=sep_tok, nan_tok=nan_tok, ) elif task == "data_imputation": train_file = data_dir_p / "train.csv" valid_file = data_dir_p / "valid.csv" test_file = data_dir_p / "test.csv" label_col = constants.IMPUTE_COLS[data_dir] read_data_func = partial( read_imputation_single, impute_col=label_col, cols_to_drop=cols_to_drop, col_renaming=col_renaming, add_instruction=add_instruction, instruction=instruction, suffix=suffix, sep_tok=sep_tok, nan_tok=nan_tok, ) elif task == "error_detection_spelling": train_file = data_dir_p / "train.csv" valid_file = data_dir_p / "valid.csv" test_file = data_dir_p / "test.csv" table_file = data_dir_p / "table.csv" table = pd.read_csv(table_file) label_col = "is_clean" read_data_func = partial( read_error_detection_single, table=table, cols_to_drop=cols_to_drop, col_renaming=col_renaming, add_instruction=add_instruction, instruction=instruction, suffix=suffix, sep_tok=sep_tok, nan_tok=nan_tok, spelling=True, ) elif task == "error_detection": train_file = data_dir_p / "train.csv" valid_file = data_dir_p / "valid.csv" test_file = data_dir_p / "test.csv" table_file = data_dir_p / "table.csv" table = pd.read_csv(table_file) label_col = "is_clean" read_data_func = partial( read_error_detection_single, table=table, cols_to_drop=cols_to_drop, col_renaming=col_renaming, add_instruction=add_instruction, instruction=instruction, suffix=suffix, sep_tok=sep_tok, nan_tok=nan_tok, spelling=False, ) elif task == "schema_matching": train_file = data_dir_p / "train.csv" valid_file = data_dir_p / "valid.csv" test_file = data_dir_p / "test.csv" table_file = data_dir_p / "table.csv" label_col = "label" table = pd.read_csv(table_file) read_data_func = partial( read_schema_match_single, table=table, cols_to_drop=cols_to_drop, col_renaming=col_renaming, add_instruction=add_instruction, instruction=instruction, suffix=suffix, sep_tok=sep_tok, nan_tok=nan_tok, ) else: raise ValueError(f"Task {task} not recognized.") data_files_sep["train"] = read_data_func(train_file) # Read validation if valid_file.exists(): data_files_sep["validation"] = read_data_func(valid_file) # Read test if test_file.exists(): data_files_sep["test"] = read_data_func(test_file) return data_files_sep, label_col The provided code snippet includes necessary dependencies for implementing the `read_data` function. Write a Python function `def read_data( data_dir: str, class_balanced: bool = False, add_instruction: bool = False, task_instruction_idx: int = 0, max_train_samples: int = -1, max_train_percent: float = -1, sep_tok: str = ".", nan_tok: str = "nan", )` to solve the following problem: Read in data where each directory is unique for a task. Here is the function: def read_data( data_dir: str, class_balanced: bool = False, add_instruction: bool = False, task_instruction_idx: int = 0, max_train_samples: int = -1, max_train_percent: float = -1, sep_tok: str = ".", nan_tok: str = "nan", ): """Read in data where each directory is unique for a task.""" data_files_sep, label_col = read_raw_data( data_dir=data_dir, add_instruction=add_instruction, task_instruction_idx=task_instruction_idx, sep_tok=sep_tok, nan_tok=nan_tok, ) task = constants.DATA2TASK[data_dir] # Don't class balance on open ended classificiation tasks if class_balanced and task != "data_imputation": # Class balance sample the train data label_cnts = data_files_sep["train"].groupby(label_col).count() sample_per_class = label_cnts.min()["text"] logger.info(f"Class balanced: train sample per class: {sample_per_class}") data_files_sep["train"] = ( data_files_sep["train"] .groupby(label_col, group_keys=False) .apply(lambda x: x.sample(sample_per_class, random_state=42)) ) # Shuffle train data data_files_sep["train"] = ( data_files_sep["train"].sample(frac=1, random_state=42).reset_index(drop=True) ) if max_train_samples > 0: orig_train_len = len(data_files_sep["train"]) if max_train_samples > 1.0: raise ValueError("max_train_samples must be between 0 and 1") max_examples = int(max_train_samples * orig_train_len) data_files_sep["train"] = data_files_sep["train"].iloc[:max_examples] logger.info( f"Length of {data_dir} train is " f"{data_files_sep['train'].shape[0]} from {orig_train_len}" ) return data_files_sep
Read in data where each directory is unique for a task.
10,177
import torch import torch.distributed as dist _COMM_DEVICE = None _PIPELINE_PARALLEL_PRED_GROUP = None _PIPELINE_PARALLEL_SUCC_GROUP = None def suppress_output(rank): """Suppress printing on the current device. Force printing with `force=True`.""" import builtins as __builtin__ builtin_print = __builtin__.print def print(*args, **kwargs): force = kwargs.pop('force', True) if force: builtin_print("rank #%d:" % rank, *args, **kwargs, flush=True) elif rank == 0: builtin_print(*args, **kwargs) __builtin__.print = print def initialize_distributed(head_ip, port, world_size, rank, local_rank, comm_device): print(f'Initializing distributed environment at {head_ip}:{port}, ' f'world_size={world_size}, rank={rank}, local_rank={local_rank}.') # Initialize distributed environment torch.cuda.set_device(local_rank) distributed_init_method = f'tcp://{head_ip}:{port}' global _COMM_DEVICE _COMM_DEVICE = comm_device if comm_device == 'cpu': backend = 'gloo' elif comm_device == 'gpu': backend = 'nccl' else: raise ValueError(f'Unknown comm_device: {comm_device}') dist.init_process_group(backend=backend, init_method=distributed_init_method, world_size=world_size, rank=rank) # Create groups for pipeline parallelism global _PIPELINE_PARALLEL_PRED_GROUP, _PIPELINE_PARALLEL_SUCC_GROUP if world_size > 1: for pred in range(world_size): succ = (pred + 1) % world_size group = dist.new_group([pred, succ]) if pred == rank: _PIPELINE_PARALLEL_PRED_GROUP = group if succ == rank: _PIPELINE_PARALLEL_SUCC_GROUP = group suppress_output(rank) print("Finished initializing distributed environment")
null
10,178
import torch import torch.distributed as dist _PIPELINE_PARALLEL_PRED_GROUP = None def get_pipeline_parallel_pred_group(): return _PIPELINE_PARALLEL_PRED_GROUP
null
10,179
import torch import torch.distributed as dist _PIPELINE_PARALLEL_SUCC_GROUP = None def get_pipeline_parallel_succ_group(): return _PIPELINE_PARALLEL_SUCC_GROUP
null
10,180
import torch import torch.distributed as dist _COMM_DEVICE = None def get_comm_device(): return _COMM_DEVICE
null
10,181
import argparse import dataclasses import os import pickle import time from typing import Union, List, Optional import numpy as np from tqdm import tqdm import torch from transformers import AutoTokenizer from flexgen.compression import CompressionConfig from flexgen.opt_config import OptConfig, get_opt_config, download_opt_weights from flexgen.pytorch_backend import (TorchDevice, TorchDisk, TorchLink, TorchMixedDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.timer import timers from flexgen.utils import (Task, ExecutionEnv, GB, T, ValueHolder, array_1d, array_2d, array_3d, str2bool, project_decode_latency, torch_mem_stats, torch_dtype_to_np_dtype, write_benchmark_log, read_benchmark_log) DUMMY_WEIGHT = "_DUMMY_" def get_choice(cur_percent, percents, choices): torch_dtype_to_np_dtype = { torch.float16: np.float16, torch.float32: np.float32, torch.uint8: np.uint8, torch.int8: np.int8, torch.int32: np.int32, torch.int64: np.int64, torch.bool: bool, } def init_weight_list(weight_specs, policy, env): dev_percents = [policy.w_disk_percent, policy.w_cpu_percent, policy.w_gpu_percent] dev_choices = [env.disk, env.cpu, env.gpu] sizes = [np.prod(spec[0]) for spec in weight_specs] sizes_cumsum = np.cumsum(sizes) ret = [] for i in range(len(weight_specs)): mid_percent = (sizes_cumsum[i] - sizes[i] / 2) / sizes_cumsum[-1] home = get_choice(mid_percent * 100, dev_percents, dev_choices) shape, dtype, filename = weight_specs[i] if len(shape) < 2: pin_memory = True compress = False else: pin_memory = policy.pin_weight compress = policy.compress_weight if not compress: weight = home.allocate(shape, dtype, pin_memory=pin_memory) if DUMMY_WEIGHT not in filename: weight.load_from_np_file(weight_specs[i][2]) else: weight.load_from_np(np.ones(shape, dtype)) #weight.load_from_np(np.random.rand(*shape).astype(dtype)) else: weight = home.compressed_device.allocate( shape, dtype, policy.comp_weight_config, pin_memory=pin_memory) if DUMMY_WEIGHT not in filename: weight.load_from_np_file(weight_specs[i][2]) else: for i in range(2): x = weight.data[i] x.load_from_np(np.ones(x.shape, torch_dtype_to_np_dtype[x.dtype])) ret.append(weight) return ret
null
10,182
import argparse import dataclasses import os import pickle import time from typing import Union, List, Optional import numpy as np from tqdm import tqdm import torch from transformers import AutoTokenizer from flexgen.compression import CompressionConfig from flexgen.opt_config import OptConfig, get_opt_config, download_opt_weights from flexgen.pytorch_backend import (TorchDevice, TorchDisk, TorchLink, TorchMixedDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.timer import timers from flexgen.utils import (Task, ExecutionEnv, GB, T, ValueHolder, array_1d, array_2d, array_3d, str2bool, project_decode_latency, torch_mem_stats, torch_dtype_to_np_dtype, write_benchmark_log, read_benchmark_log) DUMMY_WEIGHT = "_DUMMY_" class Policy: gpu_batch_size: int num_gpu_batches: int # percent = a means a% w_gpu_percent: float w_cpu_percent: float cache_gpu_percent: float cache_cpu_percent: float act_gpu_percent: float act_cpu_percent: float # Whether to overlap the I/O and compute overlap: bool # Whether to separate attention and mlp as two layers sep_layer: bool # Whether to use pinned memory for weights on CPU pin_weight: bool # Whether to compute attention on CPU cpu_cache_compute: bool # Sparsity of attention weights attn_sparsity: float # Compress weights with group-wise quantization compress_weight: bool comp_weight_config: CompressionConfig # Compress KV cache with group-wise quantization compress_cache: bool comp_cache_config: CompressionConfig def w_disk_percent(self): return 100 - self.w_gpu_percent - self.w_cpu_percent def cache_disk_percent(self): return 100 - self.cache_gpu_percent - self.cache_cpu_percent def act_disk_percent(self): return 100 - self.act_gpu_percent - self.act_cpu_percent class OptLM: def __init__(self, config: Union[str, OptConfig], env: ExecutionEnv, path: str, policy: Policy): if isinstance(config, str): config = get_opt_config(config) self.config = config self.env = env self.path = path self.policy = policy self.num_gpu_batches = policy.num_gpu_batches layers = [] layers.append(InputEmbed(self.config, self.env, self.policy)) for i in range(self.config.num_hidden_layers): if policy.sep_layer: layers.append(SelfAttention(self.config, self.env, self.policy, i)) layers.append(MLP(self.config, self.env, self.policy, i)) else: layers.append(TransformerLayer(self.config, self.env, self.policy, i)) layers.append(OutputEmbed(self.config, self.env, self.policy)) self.layers = layers self.num_layers = len(layers) if self.policy.act_gpu_percent == 100: self.act_home = self.env.gpu elif self.policy.act_cpu_percent == 100: self.act_home = self.env.cpu elif self.policy.act_disk_percent == 100: self.act_home = self.env.disk else: raise NotImplementedError() # CUDA streams self.load_weight_stream = torch.cuda.Stream() self.load_cache_stream = torch.cuda.Stream() self.store_cache_stream = torch.cuda.Stream() # Intermediate tensors # The following buffers store values used # for the i-th token, j-th layer, k-th gpu batch. num_layers, num_gpu_batches = self.num_layers, self.policy.num_gpu_batches # cache[j][k] self.cache_home = array_2d(num_layers, num_gpu_batches, ValueHolder) self.cache_read_buf = array_2d(num_layers, num_gpu_batches, ValueHolder) self.cache_write_buf = array_2d(num_layers, num_gpu_batches, ValueHolder) # weight[j] self.weight_read_buf = array_1d(num_layers, ValueHolder) # attention_mask[k] self.attention_mask = array_1d(num_gpu_batches, ValueHolder) self.task = None self.init_all_weights() def set_task(self, task): self.task = task for l in self.layers: l.set_task(task) def init_weight(self, j): expanded_path = os.path.abspath(os.path.expanduser( os.path.join(self.path, f"{self.config.name}-np"))) check_path = os.path.join(expanded_path, "decoder.embed_positions.weight") if not os.path.exists(check_path) and DUMMY_WEIGHT not in check_path: download_opt_weights(self.config.name, self.path) self.layers[j].init_weight(self.weight_home[j], expanded_path) def load_weight(self, i, j, k, overlap=True): # Handle corner cases if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load from weight_home to weight_read_buf if overlap: with torch.cuda.stream(self.load_weight_stream): self.layers[j].load_weight(self.weight_home[j], self.weight_read_buf[j], k) else: self.layers[j].load_weight(self.weight_home[j], self.weight_read_buf[j], k) def delete_weight(self, j, k): if k == 0: for x in self.weight_home[j].pop(): if isinstance(x, ValueHolder): for y in x.pop(): y.delete() else: x.delete() def init_cache(self, j, k): self.layers[j].init_cache_one_gpu_batch(self.cache_home[j][k]) def load_cache(self, i, j, k, overlap=True): # Handle corner cases if i == 0: # prefill, no cache return if k == self.num_gpu_batches: k = 0 j += 1 if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load from cache_home to cache_read_buf if overlap: with torch.cuda.stream(self.load_cache_stream): self.layers[j].load_cache(self.cache_home[j][k], self.cache_read_buf[j][k], i) else: self.layers[j].load_cache(self.cache_home[j][k], self.cache_read_buf[j][k], i) def store_cache(self, i, j, k, overlap=True): # Handle corner cases if k == -1: k = self.num_gpu_batches - 1 j -= 1 if j == -1: j = self.num_layers - 1 i -= 1 if i == -1: return if i == self.task.gen_len - 1: # last token, no need to store cache self.cache_write_buf[j][k].pop() return # Store cache_write_buf to cache_home # Delete cache_write_buf if overlap: with torch.cuda.stream(self.store_cache_stream): self.layers[j].store_cache(self.cache_home[j][k], self.cache_write_buf[j][k], i) else: self.layers[j].store_cache(self.cache_home[j][k], self.cache_write_buf[j][k], i) def delete_cache(self, j, k): v = self.cache_home[j][k].pop() if v: for x in v: x.delete() def load_hidden(self, i, j, k): # Handle corner cases if k == self.num_gpu_batches: k = 0 j += 1 if j == self.num_layers: j = 0 i += 1 if i == self.execute_gen_len: return # Load to hidden states buffers dst = self.layers[j].compute if j == 0: gpu_batch_size = self.policy.gpu_batch_size left, right = k * gpu_batch_size, (k + 1) * gpu_batch_size if i == 0: # load from the input ids val = dst.allocate((gpu_batch_size, self.task.prompt_len), np.int32) val.load_from_np(self.output_ids[left:right, :self.task.prompt_len]) else: # load from the last generated token pos = self.task.prompt_len + i val = dst.allocate((gpu_batch_size, 1), np.int32) val.load_from_np(self.output_ids[left:right, pos-1:pos]) else: # load from the last layer val = self.hidden[i][j-1][k].pop().move(dst) self.hidden[i][j][k].store(val) def store_hidden(self, i, j, k): # Handle corner cases if k == -1: k = self.num_gpu_batches - 1 j -= 1 if j == -1: j = self.num_layers - 1 i -= 1 if i == -1: return # Store to hidden states buffers if j == self.num_layers - 1: # store to output gpu_batch_size = self.policy.gpu_batch_size left, right = k * gpu_batch_size, (k + 1) * gpu_batch_size ids = self.hidden[i][j][k].pop().data.detach().cpu().numpy() pos = self.task.prompt_len + i if self.task.stop: stopped = self.stopped[left:right] self.output_ids[left:right, pos:pos+1] = np.where( stopped, self.config.pad_token_id, ids) stopped[:] = np.logical_or(stopped, ids == self.task.stop) else: self.output_ids[left:right, pos:pos+1] = ids else: # move to home x = self.hidden[i][j][k] if x.val: # x may already be moved due to overlapping x.val = x.val.move(self.act_home) def compute_layer(self, i, j, k): # Update the hidden in place # Clear the weight_read_buf if it is the last gpu batch # Clear the cache_read_buf # Run layer computation self.layers[j].forward(self.hidden[i][j][k], self.cache_read_buf[j][k], self.weight_read_buf[j], self.attention_mask[k], self.cache_write_buf[j][k], i, k) def sync(self): self.env.disk.synchronize() torch.cuda.synchronize() def init_all_weights(self): self.weight_home = array_1d(self.num_layers, ValueHolder) for j in range(self.num_layers): self.init_weight(j) def delete_all_weights(self): for j in range(self.num_layers): self.delete_weight(j, 0) def update_attention_mask(self, i, k): if i > 0: mask = self.attention_mask[k] assert mask.val is not None mask.val = mask.val.device.extend_attention_mask(mask.val, [True]) return gpu_batch_size = self.policy.gpu_batch_size left = k * gpu_batch_size right = left + gpu_batch_size input_ids = self.output_ids[left:right, :self.task.prompt_len] attention_compute = (self.env.cpu if self.policy.cpu_cache_compute else self.env.gpu) val = attention_compute.allocate( (self.policy.gpu_batch_size, self.task.prompt_len), bool) val.load_from_np((input_ids != self.config.pad_token_id)) self.attention_mask[k].store(val) def generate(self, inputs: Union[np.array, List[List[int]]], max_new_tokens: int = 32, do_sample: bool = False, temperature: float = 1.0, stop: Optional[int] = None, debug_mode: Optional[str] = None, cut_gen_len: Optional[int] = None, verbose: int = 0): task = Task( inputs=inputs, prompt_len=len(inputs[0]), gen_len=max_new_tokens, cut_gen_len=cut_gen_len, do_sample=do_sample, temperature=temperature, stop=stop, ) num_layers = self.num_layers num_gpu_batches = self.num_gpu_batches gpu_batch_size = self.policy.gpu_batch_size overlap = self.policy.overlap prompt_len, gen_len = task.prompt_len, task.gen_len self.execute_gen_len = task.cut_gen_len if task.cut_gen_len else task.gen_len # Output token ids self.output_ids = np.full((len(task.inputs), prompt_len + gen_len), self.config.pad_token_id, dtype=np.int32) self.stopped = np.zeros((len(task.inputs), 1), dtype=bool) self.output_ids[:, :prompt_len] = np.asarray(task.inputs) assert gpu_batch_size * num_gpu_batches == len(task.inputs) # Intermediate tensors # The following buffers store values used # for the i-th token, j-th layer, k-th gpu batch. num_layers, num_gpu_batches = self.num_layers, self.policy.num_gpu_batches for j in range(num_layers): for k in range(num_gpu_batches): self.cache_home[j][k].clear() self.cache_read_buf[j][k].clear() self.cache_write_buf[j][k].clear() for j in range(num_layers): self.weight_read_buf[j].clear() for k in range(num_gpu_batches): self.attention_mask[k].clear() self.hidden = array_3d(gen_len, num_layers, num_gpu_batches, ValueHolder) # Init cache self.set_task(task) for j in range(num_layers): for k in range(num_gpu_batches): self.init_cache(j, k) if self.policy.cpu_cache_compute: self.env.cpu.init_attention_compute_workspace(self.config, self.task, self.policy) # Generate if debug_mode is None: if not overlap: # No overlap, easy to understand, suitable for debugging self.generation_loop_normal() else: # Overlap I/O and compute if num_gpu_batches == 1: self.generation_loop_overlap_single_batch() else: self.generation_loop_overlap_multi_batch() elif debug_mode == "fewer_batch": # Run fewer layeres and batches for debugging if num_gpu_batches == 1: self.generation_loop_debug_single_batch() else: self.generation_loop_debug_multi_batch() elif debug_mode == "breakdown": # No overlap, fewer batches, execution time breakdown self.generation_loop_debug_normal() else: raise ValueError("Invalid debug mode: {debug_mode}") # Delete cache for j in range(num_layers): for k in range(num_gpu_batches): self.delete_cache(j, k) if self.policy.cpu_cache_compute: self.env.cpu.del_attention_compute_workspace() return self.output_ids def generation_loop_normal(self): for i in range(self.execute_gen_len): timers("generate").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): for k in range(self.num_gpu_batches): self.load_weight(i, j, k, overlap=False) for k in range(self.num_gpu_batches): self.load_cache(i, j, k, overlap=False) self.load_hidden(i, j, k) self.compute_layer(i, j, k) self.store_hidden(i, j, k) self.store_cache(i, j, k, overlap=False) timers("generate").stop() def generation_loop_debug_normal(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill_total").reset() timers("decoding_gpu_batch").reset() timers("load_weight").reset() timers("load_cache_prefill").reset() timers("load_cache_decoding").reset() timers("store_cache_prefill").reset() timers("store_cache_decoding").reset() timers("compute_layer_prefill").reset() timers("compute_layer_decoding").reset() load_weight_timer = timers("load_weight") for i in range(self.execute_gen_len): if i == 0: timers("prefill_total").start() load_cache_timer = timers("load_cache_prefill") store_cache_timer = timers("store_cache_prefill") compute_layer_timer = timers("compute_layer_prefill") else: load_cache_timer = timers("load_cache_decoding") store_cache_timer = timers("store_cache_decoding") compute_layer_timer = timers("compute_layer_decoding") for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() load_weight_timer.start(self.sync) for k in range(self.num_gpu_batches): self.load_weight(i, j, k) load_weight_timer.stop(self.sync) for k in range(self.num_gpu_batches): load_cache_timer.start(self.sync) self.load_cache(i, j, k) load_cache_timer.stop(self.sync) self.load_hidden(i, j, k) compute_layer_timer.start(self.sync) self.compute_layer(i, j, k) compute_layer_timer.stop(self.sync) self.store_hidden(i, j, k) store_cache_timer.start(self.sync) self.store_cache(i, j, k) store_cache_timer.stop(self.sync) if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill_total").stop(self.sync) # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill_total").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) # Debug the costs of individual functions print(f"#layers: {self.num_layers}") print(f"#batches prefill: " f"{self.num_layers * self.num_gpu_batches}") print(f"#batches decoding: " f"{(self.task.gen_len - 1) * self.num_layers * self.num_gpu_batches}") print(f"load_weight (per-layer)" f": {np.mean(timers('load_weight').costs):.6f} s") for stage in ["prefill", "decoding"]: for func in ["load_cache", "store_cache", "compute_layer"]: name = func + "_" + stage costs = timers(name).costs print(f"{name:22s} (per-batch): {np.mean(costs):.6f} s") def generation_loop_overlap_single_batch(self): # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.sync() # Generate for i in range(self.execute_gen_len): timers("generate").start() self.update_attention_mask(i, 0) for j in range(self.num_layers): self.load_weight(i, j+1, 0) self.load_cache(i, j+1, 0) self.load_hidden(i, j, 0) self.compute_layer(i, j, 0) self.store_cache(i, j-1, 0) self.store_hidden(i, j, 0) self.sync() timers("generate").stop() if self.task.stop and np.all(self.stopped): break def generation_loop_overlap_multi_batch(self): # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.load_hidden(0, 0, 0) self.sync() # Generate for i in range(self.execute_gen_len): timers("generate").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): for k in range(self.num_gpu_batches): self.load_weight(i, j+1, k) self.load_cache(i, j, k+1) self.store_hidden(i, j, k-1) self.load_hidden(i, j, k+1) self.compute_layer(i, j, k) self.store_cache(i, j, k-1) self.sync() timers("generate").stop() # Epilogue self.store_hidden( self.execute_gen_len-1, self.num_layers-1, self.num_gpu_batches-1) def generation_loop_debug_single_batch(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill").reset() timers("decoding_gpu_batch").reset() # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.sync() # Generate for i in range(self.execute_gen_len): if i == 0: timers("prefill").start() self.update_attention_mask(i, 0) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() self.load_weight(i, j+1, 0) self.load_cache(i, j+1, 0) self.load_hidden(i, j, 0) self.compute_layer(i, j, 0) self.store_cache(i, j-1, 0) self.store_hidden(i, j, 0) self.sync() if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill").stop() # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) def generation_loop_debug_multi_batch(self): execute_num_batches = 20 batch_ct = 0 pbar = tqdm(total=execute_num_batches) timers("prefill").reset() timers("decoding_gpu_batch").reset() # Prologue for k in range(self.num_gpu_batches): self.load_weight(0, 0, k) self.load_hidden(0, 0, 0) self.sync() # Generate for i in range(self.execute_gen_len): if i == 0: timers("prefill").start() for k in range(self.num_gpu_batches): self.update_attention_mask(i, k) for j in range(self.num_layers): if i > 0: timers("decoding_gpu_batch").start() for k in range(self.num_gpu_batches): self.load_weight(i, j+1, k) self.load_cache(i, j, k+1) self.store_hidden(i, j, k-1) self.load_hidden(i, j, k+1) self.compute_layer(i, j, k) self.store_cache(i, j, k-1) self.sync() if i > 0: timers("decoding_gpu_batch").stop() pbar.update(1) batch_ct += 1 if batch_ct >= execute_num_batches: break if batch_ct >= execute_num_batches: break if i == 0: timers("prefill").stop() # Convert "decoding_gpu_batch" timer to "generate" timer batch_cost = np.mean(timers("decoding_gpu_batch").costs[10:]) for i in range(self.execute_gen_len): if i == 0: timers("generate").costs.append(timers("prefill").costs[0]) else: timers("generate").costs.append(self.num_layers * batch_cost) def __del__(self): self.delete_all_weights() def get_filename(args): model_size = args.model.split('-')[-1] percent = "" for i in range(len(args.percent)): percent += str(args.percent[i]) + "-" filename = f"fo-{model_size}-gbs{args.gpu_batch_size}-" \ f"ngbs{args.num_gpu_batches}-" \ f"prompt{args.prompt_len}-" \ f"gen{args.gen_len}-percent-{percent}" if args.cpu_cache_compute: filename += "cpu-cache" else: filename += "gpu-cache" if args.compress_weight: filename += "-compw" if args.compress_cache: filename += "-compc" return filename def get_test_inputs(prompt_len, num_prompts, tokenizer): prompts = ["Paris is the capital city of"] input_ids = tokenizer(prompts, padding="max_length", max_length=prompt_len).input_ids return (input_ids[0],) * num_prompts class CompressionConfig: """Group-wise quantization.""" num_bits: int group_size: int group_dim: int symmetric: bool enabled: bool = True def get_opt_config(name, **kwargs): if "/" in name: name = name.split("/")[1] name = name.lower() # Handle opt-iml-30b and opt-iml-max-30b if "-iml-max" in name: arch_name = name.replace("-iml-max", "") elif "-iml" in name: arch_name = name.replace("-iml", "") else: arch_name = name if arch_name == "opt-125m": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=12, n_head=12, hidden_size=768, input_dim=768, ffn_embed_dim=768 * 4, ) elif arch_name == "opt-350m": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=24, n_head=16, hidden_size=1024, input_dim=1024, ffn_embed_dim=1024 * 4, ) raise NotImplementedError("Not implemented because this model " "has a different architecture") elif arch_name == "opt-1.3b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=24, n_head=32, hidden_size=2048, input_dim=2048, ffn_embed_dim=2048 * 4, ) elif arch_name == "opt-2.7b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=32, n_head=32, hidden_size=2560, input_dim=2560, ffn_embed_dim=2560 * 4, ) elif arch_name == "opt-6.7b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=32, n_head=32, hidden_size=4096, input_dim=4096, ffn_embed_dim=4096 * 4, ) elif arch_name == "opt-13b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=40, n_head=40, hidden_size=5120, input_dim=5120, ffn_embed_dim=5120 * 4, ) elif arch_name == "opt-30b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=48, n_head=56, hidden_size=7168, input_dim=7168, ffn_embed_dim=7168 * 4, ) elif arch_name == "galactica-30b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=48, n_head=56, hidden_size=7168, input_dim=7168, ffn_embed_dim=7168 * 4, vocab_size=50000, ) elif arch_name == "opt-66b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=64, n_head=72, hidden_size=9216, input_dim=9216, ffn_embed_dim=9216 * 4, ) elif arch_name == "opt-175b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=96, n_head=96, hidden_size=12288, input_dim=12288, ffn_embed_dim=12288 * 4, ) elif arch_name == "opt-175b-stage": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=24, n_head=96, hidden_size=12288, input_dim=12288, ffn_embed_dim=12288 * 4, ) else: raise ValueError(f"Invalid model name: {name}") return dataclasses.replace(config, **kwargs) class TorchDevice: """Wrap tensor and computation APIs of a single CPU or GPU.""" def __init__(self, name, mem_capacity=None, flops=None): self.name = name self.mem_capacity = mem_capacity self.flops = flops self.dev = torch.device(name) self.device_type = DeviceType.convert(self.dev.type) self.compressed_device = TorchCompressedDevice(self) self.links = {} self.attention_compute_workspace = None self.workspace_pt = 0 if self.device_type == DeviceType.CPU: global global_cpu_device global_cpu_device = self def add_link(self, link): dst = link.b if link.a == self else link.a self.links[dst] = link def allocate(self, shape, dtype, pin_memory=None, name=None): if self.device_type == DeviceType.CPU: pin_memory = True if pin_memory is None else pin_memory else: pin_memory = False dtype = np_dtype_to_torch_dtype[dtype] data = torch.empty(shape, dtype=dtype, pin_memory=pin_memory, device=self.dev) return TorchTensor.create_from_torch(data, self, name=name) def delete(self, tensor): pass def init_attention_compute_workspace(self, config, task, policy): if self.device_type != DeviceType.CPU: return # Only CPU requires this fp32 workspace if not policy.compress_cache: b = policy.gpu_batch_size n_head = config.n_head head_dim = config.input_dim // n_head max_seq_len = task.prompt_len + task.gen_len - 1 self.attention_compute_workspace = [] self.workspace_pt = 0 # We currently separate SelfAttention and MLP as two layers, # so we only need one workspace instead of two. for i in range(1 if policy.sep_layer else 2): shape = (max_seq_len, b * n_head, head_dim) k_cache = self.allocate(shape, np.float32, pin_memory=False) v_cache = self.allocate(shape, np.float32, pin_memory=False) self.attention_compute_workspace.append((k_cache, v_cache)) else: self.compressed_device.init_attention_compute_workspace( config, task, policy) def next_attention_compute_workspace(self): self.workspace_pt = (self.workspace_pt + 1) % len( self.attention_compute_workspace) return self.attention_compute_workspace[self.workspace_pt] def del_attention_compute_workspace(self): self.attention_compute_workspace = None def gen_attention_mask(self, token_ids, pad_token_id, donate): data = token_ids.data.ne(pad_token_id) if donate[0]: token_ids.delete() return TorchTensor.create_from_torch(data, self) def extend_attention_mask(self, attention_mask, donate): bs = attention_mask.shape[0] data = torch.concat((attention_mask.data, torch.ones((bs, 1), dtype=attention_mask.dtype, device=self.dev)), dim=1) if donate[0]: attention_mask.delete() return TorchTensor.create_from_torch(data, self) def opt_input_embed(self, inputs, attention_mask, w_token, w_pos, pad_token_id, donate): # decompress weights if w_token.device.device_type == DeviceType.COMPRESSED: w_token = w_token.device.decompress(w_token) w_pos = w_pos.device.decompress(w_pos) token_ids = inputs.data mask = attention_mask.data if donate[0]: inputs.delete() if donate[1]: attention_mask.delete() # token embedding token_embed = F.embedding(token_ids, w_token.data, pad_token_id) # pos embedding positions = torch.cumsum(mask, dim=1).int() * mask + 1 # cut positions if `past_key_values_length` is > 0 past_key_values_length = mask.shape[1] - token_ids.shape[1] positions = positions[:, past_key_values_length:] pos_embed = F.embedding(positions, w_pos.data) data = token_embed + pos_embed return TorchTensor.create_from_torch(data, self) def opt_output_embed(self, inputs, w_ln, b_ln, w_token, donate, do_sample, temperature): # decompress weights if w_token.device.device_type == DeviceType.COMPRESSED: w_token = w_token.device.decompress(w_token) b, s, h = inputs.shape hidden = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) if donate[0]: inputs.delete() # output embedding logits = F.linear(hidden, w_token.data) last_token_logits = logits[:,-1,:] if do_sample and not temperature < 1e-5: probs = torch.softmax(last_token_logits / temperature, dim=-1) ids = torch.multinomial(probs, num_samples=1) else: ids = last_token_logits.argmax(dim=1, keepdim=True) return TorchTensor.create_from_torch(ids, self) def init_cache_one_gpu_batch(self, config, task, policy): num_head, hidden_size, prompt_len, gen_len, gpu_batch_size = ( config.n_head, config.input_dim, task.prompt_len, task.gen_len, policy.gpu_batch_size) shape = (prompt_len + gen_len - 1, gpu_batch_size * num_head, hidden_size // num_head) # NOTE: disable pin_memory due to high memory overhead pin_memory = False k_cache = self.allocate(shape, np.float16, pin_memory=pin_memory) v_cache = self.allocate(shape, np.float16, pin_memory=pin_memory) return k_cache, v_cache def mha(self, inputs, attention_mask, w_q, b_q, w_k, b_k, w_v, b_v, w_out, b_out, w_ln, b_ln, n_head, donate, compress_cache, comp_config): """Multi-head attention (prefill phase).""" # decompress weights if w_q.device.device_type == DeviceType.COMPRESSED: w_q = w_q.device.decompress(w_q) w_k = w_k.device.decompress(w_k) w_v = w_v.device.decompress(w_v) w_out = w_out.device.decompress(w_out) b, s, h = inputs.shape head_dim = h // n_head scaling = head_dim ** -0.5 hidden = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) # shape: (b, s, h) q = F.linear(hidden, w_q.data, bias=b_q.data) * scaling k = F.linear(hidden, w_k.data, bias=b_k.data) v = F.linear(hidden, w_v.data, bias=b_v.data) # shape: (b, s, n_head, head_dim) q = q.view(b, s, n_head, head_dim) k = k.view(b, s, n_head, head_dim) v = v.view(b, s, n_head, head_dim) # shape: (b * n_head, s, head_dim) q = q.permute(0, 2, 1, 3).reshape(b * n_head, s, head_dim) # shape: (b * n_head, head_dim, s) k = k.permute(0, 2, 3, 1).reshape(b * n_head, head_dim, s) # shape: (b * n_head, s, head_dim) v = v.permute(0, 2, 1, 3).reshape(b * n_head, s, head_dim) # shape: (b * n_head, s, s) attn_weights = torch.bmm(q, k) # shape: (b, 1, s, s) idx = torch.arange(s, device=self.dev) causal_mask = (idx <= idx.view(s, 1)).view(1, 1, s, s) mask = attention_mask.data.view(b, 1, 1, s) & causal_mask # shape: (b, n_head, s, s) attn_weights = attn_weights.view(b, n_head, s, s) attn_weights = torch.where(mask, attn_weights, -1e4) attn_weights = attn_weights.view(b * n_head, s, s) attn_weights = F.softmax(attn_weights, dim=2) # shape: (b, n_head, s, head_dim) value = torch.bmm(attn_weights, v).view(b, n_head, s, head_dim) # shape: (b, s, h) value = value.transpose(1, 2).reshape(b, s, h) value = F.linear(value, w_out.data, bias=b_out.data) value.add_(inputs.data) if donate[0]: inputs.delete() if donate[1]: attention_mask.delete() # (s, b * n_head, head_dim) k = k.permute(2, 0, 1) v = v.permute(1, 0, 2) if compress_cache: k = self.compressed_device.compress(k, comp_config) v = self.compressed_device.compress(v, comp_config) else: k = TorchTensor.create_from_torch(k, self) v = TorchTensor.create_from_torch(v, self) return TorchTensor.create_from_torch(value, self), k, v def mha_gen(self, inputs, attention_mask, w_q, b_q, w_k, b_k, w_v, b_v, w_out, b_out, w_ln, b_ln, n_head, k_cache, v_cache, donate, attn_sparsity, compress_cache, comp_config): """Multi-head attention (decoding phase).""" # decompress weights if w_q.device.device_type == DeviceType.COMPRESSED: w_q = w_q.device.decompress(w_q) w_k = w_k.device.decompress(w_k) w_v = w_v.device.decompress(w_v) w_out = w_out.device.decompress(w_out) b, tgt_s, h = inputs.shape src_s = attention_mask.shape[1] head_dim = h // n_head scaling = head_dim ** -0.5 hidden = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) # shape: (b, 1, h) q = F.linear(hidden, w_q.data, bias=b_q.data) * scaling k = F.linear(hidden, w_k.data, bias=b_k.data) v = F.linear(hidden, w_v.data, bias=b_v.data) # shape: (b, 1, n_head, head_dim) q = q.view(b, tgt_s, n_head, head_dim) k = k.view(b, tgt_s, n_head, head_dim) v = v.view(b, tgt_s, n_head, head_dim) # shape: (b * n_head, 1, head_dim) q = q.permute(0, 2, 1, 3).reshape(b * n_head, tgt_s, head_dim) # shape: (1, b * n_head, head_dim) k_new = k.permute(1, 0, 2, 3).reshape(tgt_s, b * n_head, head_dim) # shape: (1, b * n_head, head_dim) v_new = v.permute(1, 0, 2, 3).reshape(tgt_s, b * n_head, head_dim) if isinstance(k_cache, TorchTensor): if attn_sparsity >= 1.0: # Dense attention if compress_cache: # shape: (s, b * n_head, head_dim) k = k_cache.device.decompress(k_cache)[:src_s] v = v_cache.device.decompress(v_cache)[:src_s] else: # shape: (s, b * n_head, head_dim) k = k_cache.data[:src_s] v = v_cache.data[:src_s] k[src_s - 1:src_s] = k_new v[src_s - 1:src_s] = v_new # shape: (b * n_head, head_dim, s) k = k.permute(1, 2, 0).reshape(b * n_head, head_dim, src_s) # shape: (b * n_head, s, head_dim) v = v.permute(1, 0, 2).reshape(b * n_head, src_s, head_dim) if k.is_cuda: value = self._attention_value(q, k, v, attention_mask.data, b, src_s, tgt_s, n_head, head_dim) else: q = q.float().cpu() k, v = k.float(), v.float() value = self._attention_value(q, k, v, attention_mask.data, b, src_s, tgt_s, n_head, head_dim).cuda().half() else: # Sparse attention # shape: (s, b * n_head, head_dim) k = k_cache.data[:src_s] k[src_s - 1:src_s] = k_new # shape: (b * n_head, head_dim, s) k = k.permute(1, 2, 0).reshape(b * n_head, head_dim, src_s) if k.is_cuda: value = self._sparse_attention_value(q, k, v_new, v_cache, attention_mask.data, b, src_s, tgt_s, n_head, head_dim, attn_sparsity) else: q = q.float().cpu() value = self._sparse_attention_value(q, k, v_new, v_cache, attention_mask.data, b, src_s, tgt_s, n_head, head_dim, attn_sparsity).cuda().half() else: # Mixed device attention assert attn_sparsity >= 1.0 value = self._mixed_device_attention(q, k_cache, v_cache, k_new, v_new, attention_mask.data, b, src_s, tgt_s, n_head, head_dim) # shape: (b, 1, h) value = value.transpose(1, 2).view(b, tgt_s, h) value = F.linear(value, w_out.data, bias=b_out.data) value.add_(inputs.data) if donate[0]: inputs.delete() if donate[1]: attention_mask.delete() if compress_cache: if comp_config.group_dim == 0: s_ = src_s // comp_config.group_size * comp_config.group_size k_new = k[:, :, s_:].permute(2, 0, 1) v_new = v[:, s_:, :].permute(1, 0, 2) k_new = self.compressed_device.compress(k_new, comp_config) v_new = self.compressed_device.compress(v_new, comp_config) else: k_new = TorchTensor.create_from_torch(k_new, self) v_new = TorchTensor.create_from_torch(v_new, self) return TorchTensor.create_from_torch(value, self), k_new, v_new def _attention_weights(self, q, k, mask, b, src_s, n_head): # shape: (b * n_head, 1, s) attn_weights = torch.bmm(q, k) # shape: (b, 1, 1, s) mask = mask.view(b, 1, 1, src_s) # shape: (b * n_head, 1, s) attn_weights = attn_weights.view(b, n_head, 1, src_s) attn_weights = torch.where(mask, attn_weights, -1e4) attn_weights = attn_weights.view(b * n_head, 1, src_s) attn_weights = F.softmax(attn_weights, dim=2) return attn_weights def _attention_value(self, q, k, v, mask, b, src_s, tgt_s, n_head, head_dim): # shape: (b * n_head, 1, s) attn_weights = self._attention_weights(q, k, mask, b, src_s, n_head) # shape: (b, n_head, 1, head_dim) return torch.bmm(attn_weights, v).view(b, n_head, tgt_s, head_dim) def _sparse_attention_value(self, q, k, v_new, v_cache, mask, b, src_s, tgt_s, n_head, head_dim, attn_sparsity): # shape: (b * n_head, 1, s) attn_weights = self._attention_weights(q, k, mask, b, src_s, n_head) topk = int(attn_sparsity * (attn_weights.shape[2] - 1)) topk_weights, topk_indices = attn_weights[:, :, :-1].topk( topk, dim=2, sorted=False) topk_indices = topk_indices.view(b * n_head, topk).transpose(0, 1) # shape: (b * n_head, 1, topk+1) attn_weights = torch.cat([topk_weights, attn_weights[:, :, -1].unsqueeze(-1)], dim=-1) if k.is_cuda: v_home = v_cache v_buf = self.allocate((topk+1, b*n_head, head_dim), np.float16) topk_indices = topk_indices.cpu() else: (v_home, v_buf) = v_cache # shape: (s, b * n_head, head_dim) indices_src = topk_indices indices_tgt = (slice(0, indices_src.shape[0]), slice(0, v_home.shape[1])) general_copy(v_buf, indices_tgt, v_home, indices_src) v_home.device.synchronize() # shape: (topk+1, b * n_head, head_dim) v = v_buf.data[:topk+1] v[topk:topk+1] = v_new # shape: (b * n_head, topk+1, head_dim) v = v.permute(1, 0, 2).reshape(b * n_head, topk+1, head_dim) # shape: (b * n_head, 1, head_dim) return torch.bmm(attn_weights, v).view(b, n_head, tgt_s, head_dim) def _mixed_device_attention(self, q, k_cache, v_cache, k_new, v_new, mask, b, src_s, tgt_s, n_head, head_dim): # The caches are stored on both gpu and cpu. # Compute attention on gpu for caches stored on gpu. # Compute attention on cpu for caches stored on cpu. k_gpu, k_cpu = k_cache[0].data, k_cache[1].data v_gpu, v_cpu = v_cache[0].data, v_cache[1].data seg = k_gpu.shape[1] # Compute GPU part b_gpu = seg // n_head q_gpu = q[:seg] # shape: (s, b * n_head, head_dim) k_gpu = k_gpu[:src_s, :seg, :] v_gpu = v_gpu[:src_s, :seg, :] k_gpu[src_s-1:src_s, :, :] = k_new[:, :seg, :] v_gpu[src_s-1:src_s, :, :] = v_new[:, :seg, :] # shape: (b * n_head, head_dim, s) k_gpu = k_gpu.permute(1, 2, 0) # shape: (b * n_head, s, head_dim) v_gpu = v_gpu.permute(1, 0, 2) mask_gpu = mask[:b_gpu].cuda() value_gpu = self._attention_value(q_gpu, k_gpu, v_gpu, mask_gpu, b_gpu, src_s, tgt_s, n_head, head_dim) # Compute CPU Part b_cpu = b - b_gpu q_cpu = q[seg:].float().cpu() # shape: (s, b * n_head, head_dim) k_cpu = k_cpu[:src_s, seg:, :] v_cpu = v_cpu[:src_s, seg:, :] k_cpu[src_s-1:src_s, :, :] = k_new[:, seg:, :] v_cpu[src_s-1:src_s, :, :] = v_new[:, seg:, :] # shape: (b * n_head, head_dim, s) k_cpu = k_cpu.permute(1, 2, 0) # shape: (b * n_head, s, head_dim) v_cpu = v_cpu.permute(1, 0, 2) mask_cpu = mask[b_gpu:] value_cpu = self._attention_value(q_cpu, k_cpu, v_cpu, mask_cpu, b_cpu, src_s, tgt_s, n_head, head_dim) value = torch.cat([value_gpu, value_cpu.cuda().half()], dim=0) return value def mlp(self, inputs, wi, bi, wo, bo, w_ln, b_ln, donate): # decompress weights if wi.device.device_type == DeviceType.COMPRESSED: wi = wi.device.decompress(wi) wo = wo.device.decompress(wo) b, s, h = inputs.shape out = F.layer_norm(inputs.data, (h,), weight=w_ln.data, bias=b_ln.data) out = F.linear(out, wi.data, bias=bi.data) F.relu(out, inplace=True) out = F.linear(out, wo.data, bias=bo.data) out.add_(inputs.data) if donate[0]: inputs.delete() return TorchTensor.create_from_torch(out, self) def synchronize(self): torch.cuda.synchronize() def mem_stats(self): if self.device_type == DeviceType.CUDA: cur_mem = torch.cuda.memory_allocated(self.dev) peak_mem = torch.cuda.max_memory_allocated(self.dev) elif self.device_type == DeviceType.CPU: cur_mem = cpu_mem_stats() peak_mem = 0 else: raise NotImplementedError() return cur_mem, peak_mem def print_stats(self, output_file=None): torch.cuda.synchronize() cur_mem, peak_mem = self.mem_stats() if output_file is not None: with open(output_file, "w") as f: f.write(f"TorchDevice: {self.name}\n") f.write(f" cur_mem: {cur_mem/GB:.4f} GB, " f" peak_mem: {peak_mem/GB:.4f} GB\n") else: print(f"TorchDevice: {self.name}") print(f" cur_mem: {cur_mem/GB:.4f} GB, " f" peak_mem: {peak_mem/GB:.4f} GB") return cur_mem, peak_mem def __str__(self): return f"TorchDevice(name={self.name})" class TorchDisk: """Manage tensors stored on a disk.""" def __init__(self, path, mem_capacity=None, cuda_id=0, num_copy_threads=4): self.name = path self.path = os.path.abspath(os.path.expanduser(path)) self.mem_capacity = mem_capacity self.device_type = DeviceType.DISK self.compressed_device = TorchCompressedDevice(self) if os.path.exists(self.path): assert os.path.isdir(self.path) else: os.makedirs(self.path) self.links = {} # Copy threads self.copy_queue = queue.Queue() self.copy_threads = [ threading.Thread( target=copy_worker_func, args=(self.copy_queue, cuda_id) ) for _ in range(num_copy_threads) ] for t in self.copy_threads: t.start() global global_disk_device global_disk_device = self def add_link(self, link): dst = link.b if link.a == self else link.a self.links[dst] = link def allocate(self, shape, dtype, pin_memory=None, name=None): name = name or TorchTensor.next_name() path = os.path.join(self.path, name) np.lib.format.open_memmap(path, mode="w+", shape=shape, dtype=dtype) return TorchTensor(shape, np_dtype_to_torch_dtype[dtype], path, self, name=name) def delete(self, tensor): if os.path.exists(tensor.data) and tensor.delete_file: os.remove(tensor.data) def init_cache_one_gpu_batch(self, config, task, policy): num_head, hidden_size, prompt_len, gen_len, gpu_batch_size = ( config.n_head, config.input_dim, task.prompt_len, task.gen_len, policy.gpu_batch_size) shape = (prompt_len + gen_len - 1, gpu_batch_size * num_head, hidden_size // num_head) k_cache = self.allocate(shape, np.float16) v_cache = self.allocate(shape, np.float16) return k_cache, v_cache def submit_copy(self, *args): self.copy_queue.put_nowait(args) def synchronize(self): self.copy_queue.join() def close_copy_threads(self): for _ in range(len(self.copy_threads)): self.copy_queue.put_nowait(None) for t in self.copy_threads: t.join() self.copy_queue.join() self.copy_queue = None def mem_stats(self): raise NotImplementedError() def print_stats(self): raise NotImplementedError() def __del__(self): if self.copy_queue: self.close_copy_threads() class TorchMixedDevice: """Manage tensors stored on multiple physical devices.""" def __init__(self, base_devices): self.name = "mixed" self.device_type = DeviceType.MIXED self.base_devices = base_devices def allocate(self, shape, dtype, seg_lengths, pin_memory=None, name=None): assert sum(seg_lengths) == shape[SEG_DIM] assert len(seg_lengths) == len(self.base_devices) seg_points = [0] for l in seg_lengths: seg_points.append(seg_points[-1] + l) devices = self.base_devices tensors = [] for i in range(len(devices)): seg_len = seg_points[i+1] - seg_points[i] if seg_len == 0: tensors.append(None) else: seg_shape = shape[:SEG_DIM] + (seg_len,) + shape[SEG_DIM+1:] tensors.append(devices[i].allocate(seg_shape, dtype, pin_memory=pin_memory)) return TorchTensor(shape, np_dtype_to_torch_dtype[dtype], (tensors, seg_points), self, name=name) def delete(self, tensor): for x in self.tensor.data[0]: if x: x.delete() def init_cache_one_gpu_batch(self, config, task, policy): num_head, hidden_size, prompt_len, gen_len, gpu_batch_size = ( config.n_head, config.input_dim, task.prompt_len, task.gen_len, policy.gpu_batch_size) shape = (prompt_len + gen_len - 1, gpu_batch_size * num_head, hidden_size // num_head) # We have to round to a multiple of `num_head` if policy.cache_disk_percent == 0: len_gpu = int(shape[SEG_DIM] * policy.cache_gpu_percent / 100) // num_head * num_head len_cpu = shape[SEG_DIM] - len_gpu len_disk = 0 else: len_gpu = int(shape[SEG_DIM] * policy.cache_gpu_percent / 100) // num_head * num_head len_cpu = int(shape[SEG_DIM] * policy.cache_cpu_percent / 100) // num_head * num_head len_disk = shape[SEG_DIM] - len_gpu - len_cpu lens = [len_gpu, len_cpu, len_disk] pin_memory = False k_cache = self.allocate(shape, np.float16, seg_lengths=lens, pin_memory=pin_memory) v_cache = self.allocate(shape, np.float16, seg_lengths=lens, pin_memory=pin_memory) return k_cache, v_cache timers = Timers() GB = 1 << 30 class ExecutionEnv: """Hardware environment.""" gpu: Any = None cpu: Any = None disk: Any = None mixed: Any = None def create(cls, offload_dir): # fix recursive import from flexgen.pytorch_backend import TorchDevice, TorchDisk, TorchMixedDevice gpu = TorchDevice("cuda:0") cpu = TorchDevice("cpu") disk = TorchDisk(offload_dir) return cls(gpu=gpu, cpu=cpu, disk=disk, mixed=TorchMixedDevice([gpu, cpu, disk])) def close_copy_threads(self): self.disk.close_copy_threads() def project_decode_latency(costs, prompt_len, gen_len): decode_costs = costs[1:] if gen_len / prompt_len < 0.1: warmup = 2 decode_latency = (sum(decode_costs[:warmup]) + np.mean(decode_costs[warmup:]) * (gen_len - 1 - warmup)) else: warmup = 2 decode_latency = (sum(decode_costs[:warmup]) + np.mean(decode_costs[warmup:]) * (gen_len - 1 - warmup)) #assert len(decode_costs) >= 4 #warmup = 2 #xs = np.arange(warmup, len(decode_costs)) #ys = np.asarray(decode_costs[warmup:]) #curve = np.poly1d(np.polyfit(xs, ys, deg=1)) #ys_pred = [curve(x) for x in range(gen_len-1)] #decode_latency = sum(ys_pred) #print([round(x, 4) for x in decode_costs]) #print([round(x, 4) for x in ys_pred]) return decode_latency def write_benchmark_log(filename, model_size, cache_size, hidden_size, gpu_peak_mem, projected, prefill_latency, prefill_throughput, decode_latency, decode_throughput, total_latency, total_throughput): log_str = (f"model size: {model_size/GB:.3f} GB\t" f"cache size: {cache_size/GB:.3f} GB\t" f"hidden size (p): {hidden_size/GB:.3f} GB\n" f"peak gpu mem: {gpu_peak_mem / GB:.3f} GB\t" f"projected: {projected}\n" f"prefill latency: {prefill_latency:.3f} s\t" f"prefill throughput: {prefill_throughput:.3f} token/s\n" f"decode latency: {decode_latency:.3f} s\t" f"decode throughput: {decode_throughput:.3f} token/s\n" f"total latency: {total_latency:.3f} s\t" f"total throughput: {total_throughput:.3f} token/s") with open(filename, "a") as fout: fout.write(log_str + "\n") return log_str def run_flexgen(args): print(f"<run_flexgen>: args.model: {args.model}") if args.model == "facebook/galactica-30b": tokenizer = AutoTokenizer.from_pretrained("facebook/galactica-30b", padding_side="left") else: tokenizer = AutoTokenizer.from_pretrained("facebook/opt-30b", padding_side="left") num_prompts = args.num_gpu_batches * args.gpu_batch_size prompt_len, gen_len, cut_gen_len = args.prompt_len, args.gen_len, args.cut_gen_len # Task and policy warmup_inputs = get_test_inputs(32, num_prompts, tokenizer) inputs = get_test_inputs(prompt_len, num_prompts, tokenizer) gpu = TorchDevice("cuda:0") cpu = TorchDevice("cpu") disk = TorchDisk(args.offload_dir) env = ExecutionEnv(gpu=gpu, cpu=cpu, disk=disk, mixed=TorchMixedDevice([gpu, cpu, disk])) policy = Policy(args.gpu_batch_size, args.num_gpu_batches, args.percent[0], args.percent[1], args.percent[2], args.percent[3], args.percent[4], args.percent[5], args.overlap, args.sep_layer, args.pin_weight, args.cpu_cache_compute, args.attn_sparsity, args.compress_weight, CompressionConfig(num_bits=4, group_size=64, group_dim=0, symmetric=False), args.compress_cache, CompressionConfig(num_bits=4, group_size=64, group_dim=2, symmetric=False)) assert not (args.compress_cache and args.attn_sparsity < 1.0), "Not implemented" opt_config = get_opt_config(args.model) cache_size = opt_config.cache_bytes(num_prompts, prompt_len + gen_len) hidden_size = opt_config.hidden_bytes(num_prompts, prompt_len + gen_len) print(f"model size: {opt_config.model_bytes()/GB:.3f} GB, " f"cache size: {cache_size/GB:.3f} GB, " f"hidden size (prefill): {hidden_size/GB:.3f} GB") print("init weight...") model = OptLM(opt_config, env, args.path, policy) try: print("warmup - generate") output_ids = model.generate( warmup_inputs, max_new_tokens=1, verbose=args.verbose) print("benchmark - generate") timers("generate").reset() output_ids = model.generate( inputs, max_new_tokens=args.gen_len, debug_mode=args.debug_mode, cut_gen_len=cut_gen_len, verbose=args.verbose) costs = timers("generate").costs finally: env.close_copy_threads() # Log output prefill_latency = costs[0] prefill_throughput = num_prompts * prompt_len / prefill_latency if cut_gen_len: # project latency of cut_gen_len to gen_len decode_latency = project_decode_latency(costs, prompt_len, gen_len) else: decode_latency = sum(costs[1:]) decode_throughput = num_prompts * (gen_len - 1) / max(decode_latency, 1e-10) num_generated_tokens = num_prompts * gen_len total_latency = prefill_latency + decode_latency total_throughput = num_generated_tokens / total_latency _, gpu_peak_mem = gpu.mem_stats() _, cpu_peak_mem = cpu.mem_stats() if DUMMY_WEIGHT not in args.path: outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True) show_str = "Outputs:\n" + 70 * '-' + "\n" for i in [0, len(outputs)-1]: show_str += f"{i}: {outputs[i]}\n" show_str += "-" * 70 + "\n" if args.verbose >= 2: print(show_str) gpu.print_stats() cpu.print_stats() projected = bool(args.debug_mode or cut_gen_len) if args.log_file == "auto": filename = get_filename(args) + ".log" else: filename = args.log_file log_str = write_benchmark_log(filename, opt_config.model_bytes(), cache_size, hidden_size, gpu_peak_mem, projected, prefill_latency, prefill_throughput, decode_latency, decode_throughput, total_latency, total_throughput) if args.verbose >= 1: print(log_str)
null
10,183
import argparse import dataclasses import os import pickle import time from typing import Union, List, Optional import numpy as np from tqdm import tqdm import torch from transformers import AutoTokenizer from flexgen.compression import CompressionConfig from flexgen.opt_config import OptConfig, get_opt_config, download_opt_weights from flexgen.pytorch_backend import (TorchDevice, TorchDisk, TorchLink, TorchMixedDevice, DeviceType, general_copy, fix_recursive_import) from flexgen.timer import timers from flexgen.utils import (Task, ExecutionEnv, GB, T, ValueHolder, array_1d, array_2d, array_3d, str2bool, project_decode_latency, torch_mem_stats, torch_dtype_to_np_dtype, write_benchmark_log, read_benchmark_log) def str2bool(v): if isinstance(v, bool): return v if v.lower() in ('yes', 'true', 't', 'y', '1'): return True elif v.lower() in ('no', 'false', 'f', 'n', '0'): return False else: raise argparse.ArgumentTypeError('Boolean value expected.') def add_parser_arguments(parser): parser.add_argument("--model", type=str, default="facebook/opt-6.7b", help="The model name.") parser.add_argument("--path", type=str, default="~/opt_weights", help="The path to the model weights. If there are no cached weights, " "FlexGen will automatically download them from HuggingFace.") parser.add_argument("--offload-dir", type=str, default="~/flexgen_offload_dir", help="The directory to offload tensors. ") parser.add_argument("--prompt-len", type=int, default=512) parser.add_argument("--gen-len", type=int, default=32) parser.add_argument("--cut-gen-len", type=int, help="Cut generation length for fast debugging.") parser.add_argument("--debug-mode", type=str, choices=["fewer_batch", "breakdown"]) parser.add_argument("--gpu-batch-size", type=int, default=4) parser.add_argument("--num-gpu-batches", type=int, default=1) parser.add_argument("--percent", nargs="+", type=int, default=[100, 0, 100, 0, 100, 0], help="Six numbers. They are " "the percentage of weight on GPU, " "the percentage of weight on CPU, " "the percentage of attention cache on GPU, " "the percentage of attention cache on CPU, " "the percentage of activations on GPU, " "the percentage of activations on CPU") parser.add_argument("--sep-layer", type=str2bool, nargs='?', const=True, default=True) parser.add_argument("--pin-weight", type=str2bool, nargs="?", const=True, default=True) parser.add_argument("--cpu-cache-compute", action="store_true") parser.add_argument("--attn-sparsity", type=float, default=1.0) parser.add_argument("--compress-weight", action="store_true", help="Whether to compress weight.") parser.add_argument("--compress-cache", action="store_true", help="Whether to compress cache.") parser.add_argument("--log-file", type=str, default="auto") parser.add_argument("--no-log", action="store_true") parser.add_argument("--verbose", type=int, default=2) parser.add_argument("--overlap", type=str2bool, nargs='?', const=True, default=True)
null
10,184
import argparse from itertools import count import os import pickle import traceback from typing import Union, List, Optional import numpy as np import torch import torch.distributed as dist from transformers import AutoTokenizer from flexgen.compression import CompressionConfig from flexgen.dist_utils import initialize_distributed from flexgen.flex_opt import (Policy, InputEmbed, OutputEmbed, SelfAttention, MLP, TransformerLayer, OptLM, get_filename, add_parser_arguments, get_test_inputs, DUMMY_WEIGHT) from flexgen.opt_config import get_opt_config from flexgen.pytorch_backend import (TorchDevice, TorchDisk, TorchLink, TorchMixedDevice, TorchTensor) from flexgen.timer import timers from flexgen.utils import (Task, ExecutionEnv, GB, T, ValueHolder, array_1d, array_2d, array_3d, array_4d, str2bool, project_decode_latency) class DistOptLM(OptLM): def __init__(self, config, env, path, policy, pipeline_rank, num_pipeline_stages, comm_device, num_inner_iterations=None, async_comm=False): def load_weight(self, b, t, i, j, k): def init_cache(self, t, j, k): def load_cache(self, t, i, j, k): def store_cache(self, t, i, j, k): def delete_cache(self, t, j, k): def load_hidden(self, b, t, i, j, k): def store_hidden(self, b, t, i, j, k): def send_hidden(self, t, i, j, k, tag=0, async_=False): def recv_hidden(self, t, i, j, k, tag=0, async_=False): def move_value_callback(): def compute_layer(self, t, i, j, k): def update_attention_mask(self, b, t, i, k): def generate(self, inputs: Union[np.array, List[List[int]]], max_new_tokens: int = 32, do_sample: bool = False, temperature: float = 1.0, stop: Optional[int] = None, debug_mode: Optional[str] = None, cut_gen_len: Optional[int] = None, verbose: int = 0): def send_recv_hidden(self, sending_job, receiving_job): def _send(): def _recv(): def generation_loop_normal(self): def generation_loop_overlap_one_batch(self): def generation_loop_overlap_multi_batch(self): def comm_test(comm_device): class CompressionConfig: DUMMY_WEIGHT = "_DUMMY_" class Policy: def w_disk_percent(self): def cache_disk_percent(self): def act_disk_percent(self): def get_filename(args): def get_test_inputs(prompt_len, num_prompts, tokenizer): def get_opt_config(name, **kwargs): class TorchTensor: def __init__(self, shape, dtype, data, device, name=None): def bytes(self): def next_name(cls): def create_from_torch(cls, data, device, name=None): def delete(self): def load_from_np(self, np_array): def load_from_np_file(self, filename): def copy(self, dst, src_indices=None): def smart_copy(self, dst, src_indices=None): def move(self, dst): def __str__(self): class TorchDevice: def __init__(self, name, mem_capacity=None, flops=None): def add_link(self, link): def allocate(self, shape, dtype, pin_memory=None, name=None): def delete(self, tensor): def init_attention_compute_workspace(self, config, task, policy): def next_attention_compute_workspace(self): def del_attention_compute_workspace(self): def gen_attention_mask(self, token_ids, pad_token_id, donate): def extend_attention_mask(self, attention_mask, donate): def opt_input_embed(self, inputs, attention_mask, w_token, w_pos, pad_token_id, donate): def opt_output_embed(self, inputs, w_ln, b_ln, w_token, donate, do_sample, temperature): def init_cache_one_gpu_batch(self, config, task, policy): def mha(self, inputs, attention_mask, w_q, b_q, w_k, b_k, w_v, b_v, w_out, b_out, w_ln, b_ln, n_head, donate, compress_cache, comp_config): def mha_gen(self, inputs, attention_mask, w_q, b_q, w_k, b_k, w_v, b_v, w_out, b_out, w_ln, b_ln, n_head, k_cache, v_cache, donate, attn_sparsity, compress_cache, comp_config): def _attention_weights(self, q, k, mask, b, src_s, n_head): def _attention_value(self, q, k, v, mask, b, src_s, tgt_s, n_head, head_dim): def _sparse_attention_value(self, q, k, v_new, v_cache, mask, b, src_s, tgt_s, n_head, head_dim, attn_sparsity): def _mixed_device_attention(self, q, k_cache, v_cache, k_new, v_new, mask, b, src_s, tgt_s, n_head, head_dim): def mlp(self, inputs, wi, bi, wo, bo, w_ln, b_ln, donate): def synchronize(self): def mem_stats(self): def print_stats(self, output_file=None): def __str__(self): class TorchDisk: def __init__(self, path, mem_capacity=None, cuda_id=0, num_copy_threads=4): def add_link(self, link): def allocate(self, shape, dtype, pin_memory=None, name=None): def delete(self, tensor): def init_cache_one_gpu_batch(self, config, task, policy): def submit_copy(self, *args): def synchronize(self): def close_copy_threads(self): def mem_stats(self): def print_stats(self): def __del__(self): class TorchMixedDevice: def __init__(self, base_devices): def allocate(self, shape, dtype, seg_lengths, pin_memory=None, name=None): def delete(self, tensor): def init_cache_one_gpu_batch(self, config, task, policy): timers = Timers() GB = 1 << 30 class ExecutionEnv: def create(cls, offload_dir): def close_copy_threads(self): def project_decode_latency(costs, prompt_len, gen_len): def run_flexgen_dist(args): t_name = args.model.replace("175b", "66b") tokenizer = AutoTokenizer.from_pretrained(t_name, padding_side="left") num_inner_iterations = args.num_inner_iterations if args.num_inner_iterations is not None else args.world_size num_prompts = args.num_gpu_batches * args.gpu_batch_size * num_inner_iterations * 1 prompt_len, gen_len, cut_gen_len = args.prompt_len, args.gen_len, args.cut_gen_len # Task and policy warmup_inputs = get_test_inputs(32, num_prompts, tokenizer) inputs = get_test_inputs(prompt_len, num_prompts, tokenizer) gpu = TorchDevice(f"cuda:{args.local_rank}") cpu = TorchDevice("cpu") disk = TorchDisk(args.offload_dir, None, args.local_rank) env = ExecutionEnv(gpu=gpu, cpu=cpu, disk=disk, mixed=TorchMixedDevice([gpu, cpu, disk])) TorchTensor.name_count = count(start=args.rank, step=args.world_size) comm_test(gpu.dev if args.comm_device == "gpu" else cpu.dev) policy = Policy(args.gpu_batch_size, args.num_gpu_batches, args.percent[0], args.percent[1], args.percent[2], args.percent[3], args.percent[4], args.percent[5], args.overlap, args.sep_layer, args.pin_weight, args.cpu_cache_compute, args.attn_sparsity, args.compress_weight, CompressionConfig(num_bits=4, group_size=64, group_dim=0, symmetric=False), args.compress_cache, CompressionConfig(num_bits=4, group_size=64, group_dim=2, symmetric=False)) assert not (args.compress_cache and args.attn_sparsity < 1.0), "Not implemented" opt_config = get_opt_config(args.model) model = DistOptLM(opt_config, env, args.path, policy, args.rank, args.world_size, args.comm_device, num_inner_iterations=num_inner_iterations, async_comm=args.async_comm) cache_size = opt_config.cache_bytes(num_prompts, prompt_len + gen_len) hidden_size = opt_config.hidden_bytes(num_prompts, prompt_len + gen_len) print(f"model size: {opt_config.model_bytes()/GB:.3f} GB, " f"cache size: {cache_size/GB:.3f} GB, " f"hidden size (prefill): {hidden_size/GB:.3f} GB") try: print("warmup - generate") output_ids = model.generate( warmup_inputs, max_new_tokens=2, verbose=args.verbose) print("benchmark - generate") for timer_name in ["generate-prompt", "generate"]: timers(timer_name).reset() output_ids = model.generate( inputs, max_new_tokens=args.gen_len, debug_mode=args.debug_mode, cut_gen_len=cut_gen_len, verbose=args.verbose) prompt_costs = timers("generate-prompt").costs generate_costs = timers("generate").costs finally: env.close_copy_threads() if args.rank != args.world_size - 1: return # Log output prefill_latency = sum(prompt_costs) prefill_throughput = num_prompts * prompt_len / prefill_latency if cut_gen_len: # project latency of cut_gen_len to gen_len costs = np.array(generate_costs).reshape(-1, cut_gen_len-1).sum(axis=0).tolist() decode_latency = project_decode_latency([None] + costs, prompt_len, gen_len) else: decode_latency = sum(generate_costs) decode_throughput = num_prompts * (gen_len - 1) / max(decode_latency, 1e-10) num_generated_tokens = num_prompts * gen_len total_latency = prefill_latency + decode_latency total_throughput = num_generated_tokens / total_latency _, gpu_peak_mem = gpu.mem_stats() _, cpu_peak_mem = cpu.mem_stats() if DUMMY_WEIGHT not in args.path: outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True) show_str = "Outputs:\n" + 70 * '-' + "\n" for i in [0, len(outputs)-1]: show_str += f"{i}: {outputs[i]}\n" show_str += "-" * 70 + "\n" print(show_str) gpu.print_stats() cpu.print_stats() projected = args.debug_mode or cut_gen_len log_str = (f"model size: {opt_config.model_bytes()/GB:.3f} GB\t" f"cache size: {cache_size/GB:.3f} GB\t" f"hidden size (prefill): {hidden_size/GB:.3f} GB\n" f"peak gpu mem: {gpu_peak_mem / GB:.3f} GB\n" f"prefill latency: {prefill_latency:.2f} s\t" f"prefill throughput: {prefill_throughput:.2f} token/s\n" f"decode latency: {decode_latency:.2f} s\t" f"decode throughput: {decode_throughput:.2f} token/s\n" f"total latency: {total_latency:.2f} s\t" f"total throughput: {total_throughput:.2f} token/s") print(log_str) if not args.no_log: if args.log_file == "auto": basename = f"rank-{args.rank}-{get_filename(args)}" log_filename = basename + ".log" else: log_filename = args.log_file with open(log_filename, "a") as fout: fout.write(log_str + "\n")
null
10,185
import argparse from itertools import count import os import pickle import traceback from typing import Union, List, Optional import numpy as np import torch import torch.distributed as dist from transformers import AutoTokenizer from flexgen.compression import CompressionConfig from flexgen.dist_utils import initialize_distributed from flexgen.flex_opt import (Policy, InputEmbed, OutputEmbed, SelfAttention, MLP, TransformerLayer, OptLM, get_filename, add_parser_arguments, get_test_inputs, DUMMY_WEIGHT) from flexgen.opt_config import get_opt_config from flexgen.pytorch_backend import (TorchDevice, TorchDisk, TorchLink, TorchMixedDevice, TorchTensor) from flexgen.timer import timers from flexgen.utils import (Task, ExecutionEnv, GB, T, ValueHolder, array_1d, array_2d, array_3d, array_4d, str2bool, project_decode_latency) def add_distributed_parser_arguments(parser): parser.add_argument('--head-ip', type=str, default=None, help='the IP address of the head node') parser.add_argument('--port', type=int, default=None, help='the port of the head node') parser.add_argument('--rank', metavar='I', type=int, default=None) parser.add_argument('--local-rank', metavar='I', type=int, default=None) parser.add_argument('--world-size', metavar='N', type=int, default=None) parser.add_argument('--use-mpi', action='store_true', default=False, help="Get distributed info from MPI") parser.add_argument('--comm-device', type=str, default='gpu', choices=['gpu', 'cpu'], help='communication through gpu nvlink or cpu memory ' 'and socket') parser.add_argument('--num-inner-iterations', metavar='I', type=int, default=None) parser.add_argument('--async-comm', action='store_true', default=False, help="Use asynchronous communication")
null
10,186
import time from argparse import ArgumentParser from statistics import mean import torch from petals import DistributedBloomConfig, DistributedBloomForCausalLM from torch.multiprocessing import Process, Event, Queue from transformers import AutoTokenizer, BloomConfig, OPTConfig def _patch_bloom_config(bloom_config: BloomConfig, opt_config: OPTConfig): bloom_config.hidden_size = opt_config.hidden_size bloom_config.n_head = opt_config.num_attention_heads bloom_config.n_layer = opt_config.num_hidden_layers bloom_config.vocab_size = opt_config.vocab_size
null
10,187
import time from argparse import ArgumentParser from statistics import mean import torch from petals import DistributedBloomConfig, DistributedBloomForCausalLM from torch.multiprocessing import Process, Event, Queue from transformers import AutoTokenizer, BloomConfig, OPTConfig def client_process( finished_warmup, can_start, config_bloom, num_micro_batches, batch_size, sequence_length, max_tokens, process_index, queue: Queue, ) -> None: torch.set_num_threads(1) torch.cuda.set_device(process_index % torch.cuda.device_count()) tokenizer = AutoTokenizer.from_pretrained("facebook/opt-30b") inputs = torch.randint(0, tokenizer.vocab_size, size=(batch_size, sequence_length), device="cuda") model = DistributedBloomForCausalLM(config_bloom) model.cuda() # warmup model.generate(inputs, max_new_tokens=1, do_sample=False) finished_warmup.set() can_start.wait() for _ in range(num_micro_batches): start = time.monotonic() model.generate(inputs, max_new_tokens=max_tokens, do_sample=False) end = time.monotonic() queue.put(end - start) def run_bench(args, sequence_length, max_tokens, config_bloom): queue = Queue() processes = [] warmup_events = [] can_start = Event() for i in range(args.num_processes): print("create process", i) warmup_event = Event() proc = Process(target=client_process, args=(warmup_event, can_start, config_bloom, args.num_micro_batches, args.batch_size, sequence_length, max_tokens, i, queue) ) proc.start() processes.append(proc) warmup_events.append(warmup_event) for event in warmup_events: event.wait() can_start.set() start = time.monotonic() for i, proc in enumerate(processes): print("join process", i) proc.join() end = time.monotonic() latencies = [] while not queue.empty(): latencies.append(queue.get()) print("total time", end - start) total_tokens = args.batch_size * args.num_micro_batches * args.num_processes * max_tokens print("total tokens", total_tokens) throughput = total_tokens / (end - start) print("throughput", throughput) latency = mean(latencies) print("average latency", latency) with open(args.output, "a") as f: print("\t".join( map(str, [args.batch_size, args.num_micro_batches, args.num_processes, sequence_length, max_tokens, throughput, latency] )), file=f)
null
10,188
import argparse import multiprocessing as mp import os import pickle import time import numpy as np from accelerate import (infer_auto_device_map, init_empty_weights, load_checkpoint_and_dispatch) from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM from transformers import OPTForCausalLM import torch from flexgen.timer import timers from flexgen.utils import (GB, project_decode_latency, write_benchmark_log) from flexgen.opt_config import (get_opt_config, disable_torch_init, disable_hf_opt_init) def realize_meta_module(module, dtype=None, device=None): for name, child in module.named_children(): realize_meta_module(child, dtype, device) keys = list(module._parameters.keys()) for k in keys: v = module._parameters[k] if v is not None: module._parameters[k] = torch.nn.Parameter( torch.empty(*v.shape, dtype=dtype or v.dtype, device=device or v.device)) keys = list(module._buffers.keys()) for k in keys: v = module._buffers[k] assert v is None
null
10,189
import argparse import multiprocessing as mp import os import pickle import time import numpy as np from accelerate import (infer_auto_device_map, init_empty_weights, load_checkpoint_and_dispatch) from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM from transformers import OPTForCausalLM import torch from flexgen.timer import timers from flexgen.utils import (GB, project_decode_latency, write_benchmark_log) from flexgen.opt_config import (get_opt_config, disable_torch_init, disable_hf_opt_init) def get_filename(model_name, batch_size, prompt_len, gen_len, cpu_offload, disk_offload, num_nodes, num_gpus_per_node, use_deepspeed): modelsize = model_name.split('-')[-1] if use_deepspeed: filename = "ds-" else: filename = "hf-" filename += f"{modelsize}-bs{batch_size}-prompt{prompt_len}-gen{gen_len}-" filename += f"n{num_nodes}x{num_gpus_per_node}-" if cpu_offload: filename += "cpu" elif disk_offload: filename += "disk" else: filename += "gpu" return filename def meta_to_cpu(container, dtype=None): if isinstance(container, torch.Tensor): return torch.empty(*container.shape, dtype=dtype or container.dtype) elif isinstance(container, tuple): return tuple(meta_to_cpu(x, dtype) for x in container) elif isinstance(container, dict): return dict((k, meta_to_cpu(v, dtype)) for k, v in container.items()) else: raise ValueError(f"Invalid type: {container}") def get_model_config(model_name): if "175b" in model_name: config = AutoConfig.from_pretrained("facebook/opt-66b") config.hidden_size = 12288 config.word_embed_proj_dim = 12288 config.ffn_dim = 12288 * 4 config.num_attention_heads = 96 config.num_hidden_layers = 96 else: config = AutoConfig.from_pretrained(model_name) return config def get_ds_opt_model(model_name, dtype, cpu_offload, disk_offload, offload_dir, dummy_weights): import deepspeed import torch.distributed as dist from transformers.deepspeed import HfDeepSpeedConfig config = get_model_config(model_name) hidden_size = config.hidden_size deepspeed.init_distributed("nccl") rank = dist.get_rank() pin_memory = bool(args.pin_memory) ds_config = { "fp16": { "enabled": dtype == torch.float16, }, "bf16": { "enabled": dtype == torch.bfloat16, }, "zero_optimization": { "stage": 3, "stage3_prefetch_bucket_size": hidden_size * hidden_size, "stage3_param_persistence_threshold": 0, }, "steps_per_print": 2000, "train_batch_size": args.batch_size, "wall_clock_breakdown": False, } if cpu_offload: ds_config["zero_optimization"]["offload_param"] = dict( device="cpu", pin_memory=pin_memory) if disk_offload: ds_config["zero_optimization"]["offload_param"] = dict( device="nvme", pin_memory=True, nvme_path=offload_dir, buffer_count=5, buffer_size=2 * GB, ) ds_config["aio"] = { "block_size": 1048576, "queue_depth": 8, "thread_count": 1, "single_submit": False, "overlap_events": True, } dschf = HfDeepSpeedConfig(ds_config) model = OPTForCausalLM.from_pretrained( dummy_weights or model_name, torch_dtype=dtype) model = model.eval() ds_engine = deepspeed.initialize(model=model, config_params=ds_config)[0] ds_engine.module.eval() model = ds_engine.module return model def get_hf_opt_model(model_name, dtype, cpu_offload, disk_offload, offload_dir, num_gpus, dummy_weights): if num_gpus == 1 and dtype != torch.int8: # Here we use a custom device_map instead of device_map == "auto" # becase we want to offload as many as possible weights out of GPU # to allow a larger batch size. if cpu_offload: # NOTE: We must put some weights on GPU. Otherwise, huggingface reports errors. device_map = { "model.decoder.embed_tokens.weight": 0, "model.decoder.embed_positions.weight": 0, "model.decoder.final_layer_norm": "cpu", "model.decoder.layers": "cpu", "lm_head.weight": 0, } elif disk_offload: device_map = { "model.decoder.embed_tokens.weight": 0, "model.decoder.embed_positions.weight": 0, "model.decoder.final_layer_norm": "disk", "model.decoder.layers": "disk", "lm_head.weight": 0, } else: device_map = None max_memory = None else: # Here we use device_map == "auto", but set a low `max_memory` threshold # becase we want to offload as many as possible weights out of GPU # to allow a larger batch size. device_map = "auto" if cpu_offload: # `max_memory` should be larger than the embedding. # We use 2GB here because the embeding of opt-175b is 1.2GB. max_memory = {k: "2GB" for k in range(num_gpus)} elif disk_offload: max_memory = {k: "2GB" for k in range(num_gpus)} else: max_memory = {k: "14GB" for k in range(num_gpus)} max_memory["cpu"] = "160GB" if dtype == torch.int8: kwargs = {"load_in_8bit": True} else: kwargs = {"torch_dtype": dtype} disable_torch_init() model = OPTForCausalLM.from_pretrained(dummy_weights or model_name, device_map=device_map, max_memory=max_memory, offload_folder=offload_dir, **kwargs) if device_map is None: model.cuda() model.eval() return model timers = Timers() def project_decode_latency(costs, prompt_len, gen_len): decode_costs = costs[1:] if gen_len / prompt_len < 0.1: warmup = 2 decode_latency = (sum(decode_costs[:warmup]) + np.mean(decode_costs[warmup:]) * (gen_len - 1 - warmup)) else: warmup = 2 decode_latency = (sum(decode_costs[:warmup]) + np.mean(decode_costs[warmup:]) * (gen_len - 1 - warmup)) #assert len(decode_costs) >= 4 #warmup = 2 #xs = np.arange(warmup, len(decode_costs)) #ys = np.asarray(decode_costs[warmup:]) #curve = np.poly1d(np.polyfit(xs, ys, deg=1)) #ys_pred = [curve(x) for x in range(gen_len-1)] #decode_latency = sum(ys_pred) #print([round(x, 4) for x in decode_costs]) #print([round(x, 4) for x in ys_pred]) return decode_latency def write_benchmark_log(filename, model_size, cache_size, hidden_size, gpu_peak_mem, projected, prefill_latency, prefill_throughput, decode_latency, decode_throughput, total_latency, total_throughput): log_str = (f"model size: {model_size/GB:.3f} GB\t" f"cache size: {cache_size/GB:.3f} GB\t" f"hidden size (p): {hidden_size/GB:.3f} GB\n" f"peak gpu mem: {gpu_peak_mem / GB:.3f} GB\t" f"projected: {projected}\n" f"prefill latency: {prefill_latency:.3f} s\t" f"prefill throughput: {prefill_throughput:.3f} token/s\n" f"decode latency: {decode_latency:.3f} s\t" f"decode throughput: {decode_throughput:.3f} token/s\n" f"total latency: {total_latency:.3f} s\t" f"total throughput: {total_throughput:.3f} token/s") with open(filename, "a") as fout: fout.write(log_str + "\n") return log_str def get_opt_config(name, **kwargs): if "/" in name: name = name.split("/")[1] name = name.lower() # Handle opt-iml-30b and opt-iml-max-30b if "-iml-max" in name: arch_name = name.replace("-iml-max", "") elif "-iml" in name: arch_name = name.replace("-iml", "") else: arch_name = name if arch_name == "opt-125m": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=12, n_head=12, hidden_size=768, input_dim=768, ffn_embed_dim=768 * 4, ) elif arch_name == "opt-350m": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=24, n_head=16, hidden_size=1024, input_dim=1024, ffn_embed_dim=1024 * 4, ) raise NotImplementedError("Not implemented because this model " "has a different architecture") elif arch_name == "opt-1.3b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=24, n_head=32, hidden_size=2048, input_dim=2048, ffn_embed_dim=2048 * 4, ) elif arch_name == "opt-2.7b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=32, n_head=32, hidden_size=2560, input_dim=2560, ffn_embed_dim=2560 * 4, ) elif arch_name == "opt-6.7b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=32, n_head=32, hidden_size=4096, input_dim=4096, ffn_embed_dim=4096 * 4, ) elif arch_name == "opt-13b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=40, n_head=40, hidden_size=5120, input_dim=5120, ffn_embed_dim=5120 * 4, ) elif arch_name == "opt-30b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=48, n_head=56, hidden_size=7168, input_dim=7168, ffn_embed_dim=7168 * 4, ) elif arch_name == "galactica-30b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=48, n_head=56, hidden_size=7168, input_dim=7168, ffn_embed_dim=7168 * 4, vocab_size=50000, ) elif arch_name == "opt-66b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=64, n_head=72, hidden_size=9216, input_dim=9216, ffn_embed_dim=9216 * 4, ) elif arch_name == "opt-175b": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=96, n_head=96, hidden_size=12288, input_dim=12288, ffn_embed_dim=12288 * 4, ) elif arch_name == "opt-175b-stage": config = OptConfig(name=name, max_seq_len=2048, num_hidden_layers=24, n_head=96, hidden_size=12288, input_dim=12288, ffn_embed_dim=12288 * 4, ) else: raise ValueError(f"Invalid model name: {name}") return dataclasses.replace(config, **kwargs) def run_generation(model_name, batch_size, prompt_len, gen_len, cut_gen_len, cpu_offload, disk_offload, offload_dir, use_int8, num_nodes, num_gpus_per_node, use_deepspeed, dummy, output_file, pkl_file, no_log, verbose): # Load tokenizer tokenizer = AutoTokenizer.from_pretrained( model_name.replace("175b", "66b"), padding_side="left") # Load model if use_int8: dtype = torch.int8 else: dtype = torch.float16 if dummy: config = get_model_config(model_name) filename = os.path.join(offload_dir, f"{model_name.replace('/', '-')}-hf-weights/") if not os.path.exists(filename): print("create dummy weights") with init_empty_weights(): model = OPTForCausalLM(config) model.save_pretrained(filename, state_dict=meta_to_cpu(model.state_dict(), torch.float16)) dummy_weights = filename else: dummy_weights = None print("load model") if use_deepspeed: model = get_ds_opt_model(model_name, dtype, cpu_offload, disk_offload, offload_dir, dummy_weights) else: model = get_hf_opt_model(model_name, dtype, cpu_offload, disk_offload, offload_dir, num_gpus_per_node, dummy_weights) # Run generation execute_gen_len = cut_gen_len if cut_gen_len else gen_len if use_deepspeed: prompts = ["Paris is the capital city of"] * (batch_size // WORLD_SIZE) else: prompts = ["Paris is the capital city of"] * batch_size input_ids = tokenizer(prompts, return_tensors="pt", padding="max_length", max_length=prompt_len).input_ids.cuda() # Warmup print("wamup") generate_kwargs_warmup = dict(max_new_tokens=1, do_sample=False) with torch.no_grad(): output_ids = model.generate(input_ids=input_ids, **generate_kwargs_warmup) # Run print("benchmark") timers("generate-forward").reset() generate_kwargs = dict(max_new_tokens=execute_gen_len, do_sample=False) with torch.no_grad(): output_ids = model.generate(input_ids=input_ids, **generate_kwargs) costs = timers("generate-forward").costs if use_deepspeed and args.local_rank != 0: return # Log output prefill_latency = costs[0] prefill_throughput = batch_size * prompt_len / prefill_latency if cut_gen_len: # project latency of cut_gen_len to gen_len decode_latency = project_decode_latency(costs, prompt_len, gen_len) else: decode_latency = sum(costs[1:]) decode_throughput = batch_size * (gen_len - 1) / max(decode_latency, 1e-10) num_generated_tokens = batch_size * gen_len total_latency = prefill_latency + decode_latency total_throughput = num_generated_tokens / total_latency gpu_peak_mem = torch.cuda.max_memory_allocated(torch.device("cuda")) out_str = "" if verbose >= 2: outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True) show_str = "Outputs:\n" + 70 * '-' + "\n" for i in [0, len(outputs)-1]: show_str += f"{i}: {outputs[i]}\n" show_str += 70 * '-' + "\n" print(show_str) # Check lengths input_lens = [len(x) for x in input_ids] output_lens = [len(x) for x in output_ids] assert all(x == prompt_len for x in input_lens) assert all(x == prompt_len + execute_gen_len for x in output_lens) if args.log_file == "auto": filename = get_filename(model_name, batch_size, prompt_len, gen_len, cpu_offload, disk_offload, num_nodes, num_gpus_per_node, use_deepspeed) + ".log" else: filename = args.log_file projected = bool(cut_gen_len) opt_config = get_opt_config(args.model) cache_size = opt_config.cache_bytes(batch_size, prompt_len + gen_len) hidden_size = opt_config.hidden_bytes(batch_size, prompt_len + gen_len) log_str = write_benchmark_log(filename, opt_config.model_bytes(), cache_size, hidden_size, gpu_peak_mem, projected, prefill_latency, prefill_throughput, decode_latency, decode_throughput, total_latency, total_throughput) if verbose >= 1: print(log_str)
null
10,190
import argparse from dataclasses import dataclass import time from flexgen.utils import run_cmd def run_huggingface(model, prompt_len, gen_len, cut_gen_len, batch_size, num_nodes, num_gpus_per_node, use_ds, cpu, disk, dummy, log_file=None, pkl_file=None): assert num_nodes == 1 if use_ds: cmd = f"deepspeed --num_gpus {num_gpus_per_node} hf_opt.py " else: cmd = f"python hf_opt.py --num-gpus {num_gpus_per_node} " cmd += (f"--model {model} " f"--prompt-len {prompt_len} --gen-len {gen_len} " f"--batch-size {batch_size} ") if cut_gen_len: cmd += f"--cut-gen-len {cut_gen_len} " if cpu: cmd += "--cpu " if disk: cmd += "--disk " if dummy: cmd += "--dummy " if log_file is not None: cmd += f"--log-file {log_file} " if pkl_file is not None: cmd += f"--pkl-file {pkl_file} " run_cmd(cmd) def bench_one_case(case): if case.model == "facebook/opt-6.7b": cut_gen_len = None else: cut_gen_len = 5 dummy = True if case.device == "gpu": cpu = disk = False elif case.device == "cpu": cpu, disk = True, False elif case.device == "disk": cpu, disk = False, True use_deepspeed = case.library == "ds" run_huggingface(case.model, case.prompt_len, case.gen_len, cut_gen_len, case.batch_size, case.num_nodes, case.num_gpus_per_node, use_ds=use_deepspeed, cpu=cpu, disk=disk, dummy=dummy)
null
10,191
import os import sys import subprocess from setuptools import setup, find_packages from setuptools.command import egg_info import time from op_builder import ALL_OPS, get_default_compute_capabilities, OpBuilder from op_builder.builder import installed_cuda_version ERROR = f"{RED_START} [ERROR] {RED_END}" print(f"DS_BUILD_OPS={BUILD_OP_DEFAULT}") print(f'Install Ops={install_ops}') print(f"version={version_str}, git_hash={git_hash}, git_branch={git_branch}") print(f'install_requires={install_requires}') print(f'compatible_ops={compatible_ops}') print(f'ext_modules={ext_modules}') print(f'deepspeed build time = {end_time - start_time} secs') def abort(msg): print(f"{ERROR} {msg}") assert False, msg
null
10,192
import os import sys import subprocess from setuptools import setup, find_packages from setuptools.command import egg_info import time from op_builder import ALL_OPS, get_default_compute_capabilities, OpBuilder from op_builder.builder import installed_cuda_version with open('deepspeed/git_version_info_installed.py', 'w') as fd: fd.write(f"version='{version_str}'\n") fd.write(f"git_hash='{git_hash}'\n") fd.write(f"git_branch='{git_branch}'\n") fd.write(f"installed_ops={install_ops}\n") fd.write(f"compatible_ops={compatible_ops}\n") fd.write(f"torch_info={torch_info}\n") with open(os.path.join(thisdir, 'README.md'), encoding='utf-8') as fin: readme_text = fin.read() def fetch_requirements(path): with open(path, 'r') as fd: return [r.strip() for r in fd.readlines()]
null
10,193
import os import sys import subprocess from setuptools import setup, find_packages from setuptools.command import egg_info import time from op_builder import ALL_OPS, get_default_compute_capabilities, OpBuilder from op_builder.builder import installed_cuda_version if sys.platform == "win32": # This creates a symbolic links on Windows. # It needs Administrator privilege to create symlinks on Windows. create_dir_symlink('..\\..\\csrc', '.\\deepspeed\\ops\\csrc') create_dir_symlink('..\\..\\op_builder', '.\\deepspeed\\ops\\op_builder') egg_info.manifest_maker.template = 'MANIFEST_win.in' def command_exists(cmd): if sys.platform == "win32": result = subprocess.Popen(f'{cmd}', stdout=subprocess.PIPE, shell=True) return result.wait() == 1 else: result = subprocess.Popen(f'type {cmd}', stdout=subprocess.PIPE, shell=True) return result.wait() == 0
null
10,194
import os import sys import subprocess from setuptools import setup, find_packages from setuptools.command import egg_info import time from op_builder import ALL_OPS, get_default_compute_capabilities, OpBuilder from op_builder.builder import installed_cuda_version BUILD_OP_DEFAULT = int(os.environ.get('DS_BUILD_OPS', BUILD_OP_PLATFORM)) if BUILD_OP_DEFAULT: assert torch_available, "Unable to pre-compile ops without torch installed. Please install torch before attempting to pre-compile ops." def op_envvar(op_name): if 'DS_BUILD_STRING' in os.environ: # Build string env specified, probably building for distribution with open('build.txt', 'w') as fd: fd.write(os.environ.get('DS_BUILD_STRING')) version_str += os.environ.get('DS_BUILD_STRING') elif os.path.isfile('build.txt'): # build.txt exists, probably installing from distribution with open('build.txt', 'r') as fd: version_str += fd.read().strip() else: # None of the above, probably installing from source version_str += f'+{git_hash}' def op_enabled(op_name): env_var = op_envvar(op_name) return int(os.environ.get(env_var, BUILD_OP_DEFAULT))
null
10,195
import os import sys import subprocess from setuptools import setup, find_packages from setuptools.command import egg_info import time from op_builder import ALL_OPS, get_default_compute_capabilities, OpBuilder from op_builder.builder import installed_cuda_version if 'DS_BUILD_STRING' in os.environ: # Build string env specified, probably building for distribution with open('build.txt', 'w') as fd: fd.write(os.environ.get('DS_BUILD_STRING')) version_str += os.environ.get('DS_BUILD_STRING') elif os.path.isfile('build.txt'): # build.txt exists, probably installing from distribution with open('build.txt', 'r') as fd: version_str += fd.read().strip() else: # None of the above, probably installing from source version_str += f'+{git_hash}' def create_dir_symlink(src, dest): if not os.path.islink(dest): if os.path.exists(dest): os.remove(dest) assert not os.path.exists(dest) os.symlink(src, dest)
null
10,196
import os from .constants import (MODEL_FILE_PREFIX, MODEL_FILE_SUFFIX, OPTIM_FILE_SUFFIX, ZERO_FILE_PREFIX) MODEL_FILE_PREFIX = 'mp_rank_' MODEL_FILE_SUFFIX = '_model_states.pt' def get_model_ckpt_name_for_rank(base_folder, mp_rank_str): ckpt_name = os.path.join( base_folder, MODEL_FILE_PREFIX + mp_rank_str + MODEL_FILE_SUFFIX, ) return ckpt_name
null
10,197
import os from .constants import (MODEL_FILE_PREFIX, MODEL_FILE_SUFFIX, OPTIM_FILE_SUFFIX, ZERO_FILE_PREFIX) MODEL_FILE_PREFIX = 'mp_rank_' ZERO_FILE_PREFIX = 'zero_pp_rank_' OPTIM_FILE_SUFFIX = '_optim_states.pt' def get_zero_ckpt_name_for_rank(base_folder, dp_rank, mp_rank): zero_prefix = f'{ZERO_FILE_PREFIX}{dp_rank}' mp_rank_string = f'_{MODEL_FILE_PREFIX}{mp_rank:02d}' zero_ckpt_name = os.path.join( base_folder, zero_prefix + mp_rank_string + OPTIM_FILE_SUFFIX, ) return zero_ckpt_name
null
10,198
import os from .constants import (MODEL_FILE_PREFIX, MODEL_FILE_SUFFIX, OPTIM_FILE_SUFFIX, ZERO_FILE_PREFIX) MODEL_FILE_SUFFIX = '_model_states.pt' def get_layer_ckpt_name_for_rank(base_folder, layer_id, tp_rank): ckpt_file = f'{layer_id}-model_{tp_rank:02d}{MODEL_FILE_SUFFIX}' ckpt_path = os.path.join(base_folder, ckpt_file) return ckpt_path
null
10,199
from .reshape_utils import partition_data class meg_2d_parallel_map(object): def __init__(self, pp_degree, tp_degree): self.pp_degree = pp_degree self.tp_degree = tp_degree self.map = {} def simple_init(self): self.map = { self._make_key(i // self.tp_degree, i % self.tp_degree): [i] for i in range(self.pp_degree * self.tp_degree) } def add_data(self, pp_index, tp_index, data): self._validate_indices(pp_index, tp_index) assert type(data) is list key = self._make_key(pp_index, tp_index) if not key in self.map.keys(): self.map[key] = [] self.map[key] += data def get_data(self, pp_index=None, tp_index=None): self._validate_indices(pp_index, tp_index) pp_indices = list(range(self.pp_degree)) if pp_index is None else [pp_index] tp_indices = list(range(self.tp_degree)) if tp_index is None else [tp_index] result = [] for i in pp_indices: for j in tp_indices: result += self.map[self._make_key(i, j)] return result def print_data(self, tag): print(f'{tag}') for key, value in self.map.items(): print(f'{key} = {value}') def _validate_indices(self, pp_index, tp_index): assert pp_index is None or pp_index < self.pp_degree assert tp_index is None or tp_index < self.tp_degree def _make_key(self, i, j): return f'{i},{j}' def _reshape_tp_dimension(old_2d_map, new_tp_degree): old_pp_degree = old_2d_map.pp_degree new_2d_map = meg_2d_parallel_map(old_pp_degree, new_tp_degree) for i in range(old_pp_degree): ranks_for_pp_index = old_2d_map.get_data(pp_index=i, tp_index=None) split_ranks = partition_data(ranks_for_pp_index, new_tp_degree) for j in range(new_tp_degree): new_2d_map.add_data(i, j, split_ranks[j]) return new_2d_map def _reshape_pp_dimension(old_2d_map, new_pp_degree): old_tp_degree = old_2d_map.tp_degree new_2d_map = meg_2d_parallel_map(new_pp_degree, old_tp_degree) for i in range(old_tp_degree): ranks_for_tp_index = old_2d_map.get_data(pp_index=None, tp_index=i) split_ranks = partition_data(ranks_for_tp_index, new_pp_degree) for j in range(new_pp_degree): new_2d_map.add_data(j, i, split_ranks[j]) return new_2d_map def reshape_meg_2d_parallel(old_pp_degree, old_tp_degree, new_pp_degree, new_tp_degree, verbose=False): assert new_pp_degree <= old_pp_degree assert new_tp_degree <= old_tp_degree old_2d_map = meg_2d_parallel_map(old_pp_degree, old_tp_degree) old_2d_map.simple_init() if verbose: old_2d_map.print_data(f'original_2d_map:') if old_tp_degree != new_tp_degree: new_tp_map = _reshape_tp_dimension(old_2d_map, new_tp_degree) else: new_tp_map = old_2d_map if verbose: new_tp_map.print_data(f'after_tp_reshape:') if old_pp_degree != new_pp_degree: final_map = _reshape_pp_dimension(new_tp_map, new_pp_degree) else: final_map = new_tp_map if verbose: final_map.print_data(f'final_2d_map:') return final_map
null
10,200
from .reshape_utils import partition_data def get_mpu_ranks(tp_size=1, pp_size=1, dp_size=1, virtual_pp_size=None): """ Initialize model data parallel groups. Arguments: tp_size: number of GPUs used to parallelize model tensor. pp_size: number of GPUs used to parallelize model pipeline. dp_size: number of GPUs used to parallelize model data. Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize the model pipeline. The present function will create 8 tensor model-parallel groups, 4 pipeline model-parallel groups and 8 data-parallel groups as: 8 data_parallel groups: [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15] 8 tensor model-parallel groups: [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15] 4 pipeline model-parallel groups: [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15] Note that for efficiency, the caller should make sure adjacent ranks are on the same DGX box. For example if we are using 2 DGX-1 boxes with a total of 16 GPUs, rank 0 to 7 belong to the first box and ranks 8 to 15 belong to the second box. """ world_size = tp_size * pp_size * dp_size print(f"\n\n*** tp={tp_size}, pp={pp_size}, dp={dp_size}, world={world_size}") tensor_model_parallel_size = min(tp_size, world_size) pipeline_model_parallel_size = min(pp_size, world_size) data_parallel_size = world_size // (tensor_model_parallel_size * pipeline_model_parallel_size) num_tensor_model_parallel_groups = world_size // tensor_model_parallel_size num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size num_data_parallel_groups = world_size // data_parallel_size # Build the data-parallel groups. all_dp_group_ranks = [] for i in range(pipeline_model_parallel_size): start_rank = i * num_pipeline_model_parallel_groups end_rank = (i + 1) * num_pipeline_model_parallel_groups for j in range(tensor_model_parallel_size): ranks = range(start_rank + j, end_rank, tensor_model_parallel_size) all_dp_group_ranks.append(list(ranks)) print("DP", all_dp_group_ranks) # Build the model-parallel groups. all_pp_group_ranks = [] for i in range(data_parallel_size): ranks = [ data_parallel_group_ranks[i] for data_parallel_group_ranks in all_dp_group_ranks ] all_pp_group_ranks.append(list(ranks)) print(f"PP", all_pp_group_ranks) # Build the tensor model-parallel groups. all_tp_group_ranks = [] for i in range(num_tensor_model_parallel_groups): ranks = range(i * tensor_model_parallel_size, (i + 1) * tensor_model_parallel_size) all_tp_group_ranks.append(list(ranks)) print(f"TP", all_tp_group_ranks) return all_tp_group_ranks, all_pp_group_ranks, all_dp_group_ranks # # Build the pipeline model-parallel groups and embedding groups # # (first and last rank in each pipeline model-parallel group). # for i in range(num_pipeline_model_parallel_groups): # ranks = range(i, world_size, # num_pipeline_model_parallel_groups) # print(f"EMB{i}", list(ranks)) The provided code snippet includes necessary dependencies for implementing the `reshape` function. Write a Python function `def reshape(src, tgt)` to solve the following problem: reshape([tp_size_src, pp_size_src, dp_size_src], [tp_size_tgt, pp_size_tgt, dp_size_tgt]) Here is the function: def reshape(src, tgt): """ reshape([tp_size_src, pp_size_src, dp_size_src], [tp_size_tgt, pp_size_tgt, dp_size_tgt]) """ print(f"\n\n*** Reshaping: {src} => {tgt}") tp_size_src, pp_size_src, dp_size_src = src tp_size_tgt, pp_size_tgt, dp_size_tgt = tgt tp_ranks1, pp_ranks1, dp_ranks1 = get_mpu_ranks(tp_size=tp_size_src, pp_size=pp_size_src, dp_size=dp_size_src) tp_ranks2, pp_ranks2, dp_ranks2 = get_mpu_ranks(tp_size=tp_size_tgt, pp_size=pp_size_src, dp_size=dp_size_src) tp_ranks3, pp_ranks3, dp_ranks3 = get_mpu_ranks(tp_size=tp_size_tgt, pp_size=pp_size_tgt, dp_size=dp_size_src) # handle tp contraction first print("\n*** TP contraction:") for i, r in enumerate(tp_ranks1): print(f'{tp_ranks1[i]} => {tp_ranks2[i]}') # handle pp contraction next print("\n*** PP contraction:") for i, r in enumerate(pp_ranks1): print(f'{pp_ranks2[i]} => {pp_ranks3[i]}')
reshape([tp_size_src, pp_size_src, dp_size_src], [tp_size_tgt, pp_size_tgt, dp_size_tgt])
10,201
import os import torch from collections import OrderedDict from .constants import (ZERO_FILE_PREFIX, FP16_ZERO_FILE_PREFIX, BF16_ZERO_FILE_PREFIX) def basic_folder_validation(dir): assert os.path.exists(dir), f'{dir} path does not exist' assert os.path.isdir(dir), f'{dir} is not a folder'
null
10,202
import os import torch from collections import OrderedDict from .constants import (ZERO_FILE_PREFIX, FP16_ZERO_FILE_PREFIX, BF16_ZERO_FILE_PREFIX) def validate_files(file_list): for file in file_list: if not os.path.isfile(file): print(f'Error: {file} is not existent')
null
10,203
import os import torch import types from .constants import (FP32_WEIGHT_KEY, PARAM, VOCAB_DIVISIBILITY_PADDING_TENSOR, CAT_DIM) def load_hp_checkpoint_state(self, folder, tp_rank, tp_world_size): hp_mapping = self._hp_mapping optim_state_keys = hp_mapping.get_optim_state_keys() hp_keys = [FP32_WEIGHT_KEY] + optim_state_keys checkpoint_files = {key: os.path.join(folder, f"{key}.pt") for key in hp_keys} for file in checkpoint_files.values(): assert os.path.isfile(file), f'{file} is not a valid file' for key in hp_keys: ckpt_file = checkpoint_files[key] ckpt_dict = torch.load(ckpt_file) full_hp_param = ckpt_dict[PARAM] # need to deal with slices that were averaged. # the opposite of averaging here becomes an exact copy of the first slice # I thought of 2 ways: # implementation a. find a way for a client to pass a dict with patterns # if any(re.search(pattern, folder) for pattern in WEIGHTS_TO_AVERAGE_PATTERNS): # tp_rank = 0 # tp_world_size = 1 # the other approach is to assume that the saved data is correct and if full_hp_param.shape == # self.shape that means we automatically copy? # implementation b. # this version requires no additional data passed from the client # if the shapes already match it must be slices that were averaged - so we just hack around those if full_hp_param.shape == self.shape: tp_rank = 0 tp_world_size = 1 # special case for word_embeddings weights which get padded differently depending on TP degree. # the converter to universal currently strips the original padding completely so the saved # weight is padding-free and we just need to add new padding depending on the target TP # degree vocab_divisibility_padding_tensor = ckpt_dict.get( VOCAB_DIVISIBILITY_PADDING_TENSOR, None) if vocab_divisibility_padding_tensor is not None: # In the absence of data passed from the user wrt new padded vocab specific to tp degree # we can again derive that data by reverse engineering the target shapes like so: padded_target_vocab_size = self.shape[0] * tp_world_size if padded_target_vocab_size > full_hp_param.shape[0]: # Need to expand padding_size = padded_target_vocab_size - full_hp_param.shape[0] # Implement the following concat in efficient way using pad #full_hp_param = torch.cat((full_hp_param, padding_tensor), 0) full_hp_param = torch.nn.functional.pad(full_hp_param, (0, 0, 0, padding_size), "constant", 0) full_hp_param[:-padding_size, :] = vocab_divisibility_padding_tensor else: # Need to shrink or keep the same full_hp_param = full_hp_param[:padded_target_vocab_size, :] full_param_numel = full_hp_param.numel() tp_slice_numel = self.numel() # if key == FP32_WEIGHT_KEY and 'word_embeddings.weight' in folder: # print_rank_0(f'{full_hp_param[:10]=}', force=True) assert full_param_numel == tp_world_size * tp_slice_numel, \ f'Loading {ckpt_file} full param numel {full_param_numel} != tensor slice numel {tp_slice_numel} * tp_world_size {tp_world_size}' dst_tensor = hp_mapping.hp_fragment if key == FP32_WEIGHT_KEY else hp_mapping.get_optim_state_fragment( key) # print(f"{full_hp_param.shape=} {full_param_numel=} {folder=}") # print(f"{dst_tensor.shape=} {dst_tensor.numel()=}{folder=}") # since when we do many to 1 on tp we cat sometimes on dim=0 and other times on dim=1 we have to do exactly the same in reverse chunk_dim = ckpt_dict.get(CAT_DIM, 0) # this performs the opposite of cat when merging TP slices tp_hp_slice = full_hp_param.chunk(tp_world_size, chunk_dim)[tp_rank] tp_hp_slice = tp_hp_slice.flatten() lp_frag_address = hp_mapping.lp_fragment_address tp_hp_fragment = tp_hp_slice.narrow(0, lp_frag_address.start, lp_frag_address.numel) assert dst_tensor.numel() == lp_frag_address.numel, \ f'Load checkpoint {key} dst_tensor numel {dst_tensor.numel()} != src numel {lp_frag_address.numel}' # print(f"{key} SHAPE: {tp_hp_slice.shape=}") # print(f"{key} SHAPE: {dst_tensor.shape=}") # print(f"{key} SHAPE: {tp_hp_fragment.shape=}") dst_tensor.data.copy_(tp_hp_fragment.data) def enable_universal_checkpoint(param_list): for param in param_list: param.load_hp_checkpoint_state = types.MethodType(load_hp_checkpoint_state, param)
null
10,204
from .reshape_utils import (get_files, get_files_with_prefix, partition_data, get_zero_files) from .constants import (MODEL_FILE_PREFIX, LAYER_FILE_PREFIX) from .reshape_meg_2d import (reshape_meg_2d_parallel, meg_2d_parallel_map) class model_3d_desc(object): def __init__(self, pp_degree=1, tp_degree=1, dp_degree=1): self.pp_degree = pp_degree self.tp_degree = tp_degree self.dp_degree = dp_degree def reshape(self, target_3d_desc, verbose=False): valid_reshape, reshape_errors = self.can_reshape(target_3d_desc) assert valid_reshape, ','.join(reshape_errors) tgt_2d_map = reshape_meg_2d_parallel(old_pp_degree=self.pp_degree, old_tp_degree=self.tp_degree, new_pp_degree=target_3d_desc.pp_degree, new_tp_degree=target_3d_desc.tp_degree, verbose=verbose) flat_3d_map = flatten_dp_dimension(meg_2d_map=tgt_2d_map, src_2d_size=self.pp_degree * self.tp_degree, dp_degree=self.dp_degree) return unflatten_dp_dimension(meg_2d_map=flat_3d_map, dp_degree=target_3d_desc.dp_degree) def get_desc(self): return f'{PP_DIM},{TP_DIM},{DP_DIM} = ({self.pp_degree}, {self.tp_degree}, {self.dp_degree})' def world_size(self): return self.pp_degree * self.tp_degree * self.dp_degree def is_valid(self, pp_index, tp_index, dp_index): err_msg = [] valid = True for index, degree, dim_name in [ (pp_index, self.pp_degree, PP_DIM), (tp_index, self.tp_degree, TP_DIM), (dp_index, self.dp_degree, DP_DIM)]: if index >= degree: valid = False err_msg.append( f'{dim_name} indexing error: index {index} >= degree {degree}') return valid, err_msg def can_reshape(self, target_3d_desc): err_msg = [] if target_3d_desc.pp_degree > self.pp_degree: err_msg.append( f'Expansion reshape not supported - {PP_DIM}: {self.pp_degree} ---> {target_3d_desc.pp_degree}' ) if target_3d_desc.tp_degree > self.tp_degree: err_msg.append( f'Expansion reshape not supported - {TP_DIM}: {self.tp_degree} ---> {target_3d_desc.tp_degree}' ) if target_3d_desc.dp_degree > self.dp_degree: err_msg.append( f'Expansion reshape not supported - {DP_DIM}: {self.dp_degree} ---> {target_3d_desc.dp_degree}' ) return len(err_msg) == 0, err_msg def get_files_with_prefix(all_files, prefix): file_list = [] for file_path in all_files: _, fname = os.path.split(file_path) if fname.startswith(prefix): file_list.append(file_path) return sorted(file_list) def get_files(dir): file_list = [] for root, _, files in os.walk(dir): for file in files: file_list.append(os.path.join(root, file)) return file_list def get_zero_files(dir): file_list = get_files(dir) for prefix in [ZERO_FILE_PREFIX, FP16_ZERO_FILE_PREFIX, BF16_ZERO_FILE_PREFIX]: zero_files = get_files_with_prefix(file_list, prefix) if len(zero_files) > 0: return zero_files return [] MODEL_FILE_PREFIX = 'mp_rank_' LAYER_FILE_PREFIX = 'layer_' def get_model_3d_descriptor(dir): file_list = get_files(dir) zero_file_list = get_zero_files(dir) num_pp0_files = len(get_files_with_prefix(file_list, f'{LAYER_FILE_PREFIX}01')) if num_pp0_files > 0: tp_degree = num_pp0_files pp_degree = len(get_files_with_prefix(file_list, MODEL_FILE_PREFIX)) // tp_degree dp_degree = max(1, len(zero_file_list) // (pp_degree * tp_degree)) else: tp_degree = len(get_files_with_prefix(file_list, MODEL_FILE_PREFIX)) dp_degree = max(1, len(zero_file_list) // tp_degree) pp_degree = 0 return model_3d_desc(pp_degree, tp_degree, dp_degree)
null
10,205
from .reshape_utils import (get_files, get_files_with_prefix, partition_data, get_zero_files) from .constants import (MODEL_FILE_PREFIX, LAYER_FILE_PREFIX) from .reshape_meg_2d import (reshape_meg_2d_parallel, meg_2d_parallel_map) class meg_2d_parallel_map(object): def __init__(self, pp_degree, tp_degree): def simple_init(self): def add_data(self, pp_index, tp_index, data): def get_data(self, pp_index=None, tp_index=None): def print_data(self, tag): def _validate_indices(self, pp_index, tp_index): def _make_key(self, i, j): def flatten_dp_dimension(meg_2d_map, src_2d_size, dp_degree): new_meg_2d_map = meg_2d_parallel_map(meg_2d_map.pp_degree, meg_2d_map.tp_degree) for pp_index in range(meg_2d_map.pp_degree): for tp_index in range(meg_2d_map.tp_degree): dp0_indices = meg_2d_map.get_data(pp_index, tp_index) for idx in dp0_indices: dpX_indices = [idx + (i * src_2d_size) for i in range(dp_degree)] new_meg_2d_map.add_data(pp_index, tp_index, dpX_indices) return new_meg_2d_map
null
10,206
from .reshape_utils import (get_files, get_files_with_prefix, partition_data, get_zero_files) from .constants import (MODEL_FILE_PREFIX, LAYER_FILE_PREFIX) from .reshape_meg_2d import (reshape_meg_2d_parallel, meg_2d_parallel_map) def partition_data(data_list, num_partitions): num_elems = len(data_list) assert num_elems % num_partitions == 0 partition_size = num_elems // num_partitions partitions_list = [ data_list[i:i + partition_size] for i in range(0, num_elems, partition_size) ] return partitions_list class meg_2d_parallel_map(object): def __init__(self, pp_degree, tp_degree): self.pp_degree = pp_degree self.tp_degree = tp_degree self.map = {} def simple_init(self): self.map = { self._make_key(i // self.tp_degree, i % self.tp_degree): [i] for i in range(self.pp_degree * self.tp_degree) } def add_data(self, pp_index, tp_index, data): self._validate_indices(pp_index, tp_index) assert type(data) is list key = self._make_key(pp_index, tp_index) if not key in self.map.keys(): self.map[key] = [] self.map[key] += data def get_data(self, pp_index=None, tp_index=None): self._validate_indices(pp_index, tp_index) pp_indices = list(range(self.pp_degree)) if pp_index is None else [pp_index] tp_indices = list(range(self.tp_degree)) if tp_index is None else [tp_index] result = [] for i in pp_indices: for j in tp_indices: result += self.map[self._make_key(i, j)] return result def print_data(self, tag): print(f'{tag}') for key, value in self.map.items(): print(f'{key} = {value}') def _validate_indices(self, pp_index, tp_index): assert pp_index is None or pp_index < self.pp_degree assert tp_index is None or tp_index < self.tp_degree def _make_key(self, i, j): return f'{i},{j}' def unflatten_dp_dimension(meg_2d_map, dp_degree): pp_degree = meg_2d_map.pp_degree tp_degree = meg_2d_map.tp_degree meg_2d_map_list = [ meg_2d_parallel_map(pp_degree=pp_degree, tp_degree=tp_degree) for _ in range(dp_degree) ] for pp_index in range(pp_degree): for tp_index in range(tp_degree): flat_dp_indices = meg_2d_map.get_data(pp_index, tp_index) partitioned_dp_indices = partition_data(flat_dp_indices, dp_degree) for dp_indices, _2d_map in zip(partitioned_dp_indices, meg_2d_map_list): _2d_map.add_data(pp_index, tp_index, dp_indices) return meg_2d_map_list
null
10,207
from collections import OrderedDict import torch import sys import os from deepspeed import comm as dist from deepspeed.runtime.constants import PIPE_REPLICATED from deepspeed.ops.op_builder import UtilsBuilder from deepspeed.runtime import ZeROOptimizer from packaging import version as pkg_version from deepspeed.git_version_info import version from deepspeed.runtime.utils import (get_global_norm_of_tensors, clip_tensors_by_global_norm, DummyOptim, align_dense_tensors, all_gather_dp_groups, bwc_tensor_model_parallel_rank, is_model_parallel_parameter, see_memory_usage) from deepspeed.utils import link_hp_params, fragment_address from deepspeed.checkpoint import enable_universal_checkpoint from deepspeed.checkpoint.constants import (DS_VERSION, PARTITION_COUNT, BASE_OPTIMIZER_STATE, SINGLE_PARTITION_OF_FP32_GROUPS, CLIP_GRAD, GROUP_PADDINGS, PARAM_SLICE_MAPPINGS) def _get_padded_tensor(src_tensor, size): if src_tensor.numel() >= size: return src_tensor padded_tensor = torch.zeros(size, dtype=src_tensor.dtype, device=src_tensor.device) slice_tensor = torch.narrow(padded_tensor, 0, 0, src_tensor.numel()) slice_tensor.data.copy_(src_tensor.data) return padded_tensor
null
10,208
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def noop_decorator(func): return func
null
10,209
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod The provided code snippet includes necessary dependencies for implementing the `ensure_directory_exists` function. Write a Python function `def ensure_directory_exists(filename)` to solve the following problem: Create the directory path to ``filename`` if it does not already exist. Args: filename (str): A file path. Here is the function: def ensure_directory_exists(filename): """Create the directory path to ``filename`` if it does not already exist. Args: filename (str): A file path. """ dirname = os.path.dirname(filename) os.makedirs(dirname, exist_ok=True)
Create the directory path to ``filename`` if it does not already exist. Args: filename (str): A file path.
10,210
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod The provided code snippet includes necessary dependencies for implementing the `set_random_seed` function. Write a Python function `def set_random_seed(seed)` to solve the following problem: Set the random seed for common PRNGs used during training: random, numpy, and torch. Args: seed (int): the seed to use Here is the function: def set_random_seed(seed): """Set the random seed for common PRNGs used during training: random, numpy, and torch. Args: seed (int): the seed to use """ import numpy import random random.seed(seed) numpy.random.seed(seed) torch.manual_seed(seed)
Set the random seed for common PRNGs used during training: random, numpy, and torch. Args: seed (int): the seed to use
10,211
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod The provided code snippet includes necessary dependencies for implementing the `copy_to_device` function. Write a Python function `def copy_to_device(item, device, criterion_func)` to solve the following problem: Return a copy of tensor on specified device. Works on individual tensors, and tensors contained/nested in lists, tuples, and dicts. Parameters: item: tensor to copy or (possibly nested) container of tensors to copy. device: target device criterion_func: Function to restrict copy operation to items meet criterion Returns: None Here is the function: def copy_to_device(item, device, criterion_func): """ Return a copy of tensor on specified device. Works on individual tensors, and tensors contained/nested in lists, tuples, and dicts. Parameters: item: tensor to copy or (possibly nested) container of tensors to copy. device: target device criterion_func: Function to restrict copy operation to items meet criterion Returns: None """ if criterion_func(item): return item.to(device) elif isinstance(item, list): return [copy_to_device(v, device, criterion_func) for v in item] elif isinstance(item, tuple): return tuple([copy_to_device(v, device, criterion_func) for v in item]) elif isinstance(item, dict): return {k: copy_to_device(v, device, criterion_func) for k, v in item.items()} else: return item
Return a copy of tensor on specified device. Works on individual tensors, and tensors contained/nested in lists, tuples, and dicts. Parameters: item: tensor to copy or (possibly nested) container of tensors to copy. device: target device criterion_func: Function to restrict copy operation to items meet criterion Returns: None
10,212
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod The provided code snippet includes necessary dependencies for implementing the `move_to_device` function. Write a Python function `def move_to_device(item, device, criterion_func)` to solve the following problem: Move tensor on to specified device by changing the storage. Works on individual tensors, and tensors contained/nested in lists, tuples, and dicts. Parameters: item: tensor to move or (possibly nested) container of tensors to move. device: target device criterion_func: Function to restrict move operation to items meet criterion Returns: None Here is the function: def move_to_device(item, device, criterion_func): """ Move tensor on to specified device by changing the storage. Works on individual tensors, and tensors contained/nested in lists, tuples, and dicts. Parameters: item: tensor to move or (possibly nested) container of tensors to move. device: target device criterion_func: Function to restrict move operation to items meet criterion Returns: None """ if criterion_func(item): device_copy = item.to(device) item.data = device_copy.data return item elif isinstance(item, list): return [move_to_device(v, device, criterion_func) for v in item] elif isinstance(item, tuple): return tuple([move_to_device(v, device, criterion_func) for v in item]) elif isinstance(item, dict): return {k: move_to_device(v, device, criterion_func) for k, v in item.items()} else: return item
Move tensor on to specified device by changing the storage. Works on individual tensors, and tensors contained/nested in lists, tuples, and dicts. Parameters: item: tensor to move or (possibly nested) container of tensors to move. device: target device criterion_func: Function to restrict move operation to items meet criterion Returns: None
10,213
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def _handle_overflow(cpu_sum, x, i): import math rank = dist.get_rank() if rank == 0: t_i = -1 for v_i, v in enumerate(x.data.contiguous().view(-1)): if not math.isfinite(float(v)): t_i = v_i break logger.info( f"rank {rank} detected overflow {cpu_sum} in tensor {i}:{t_i} shape {x.shape}" )
null
10,214
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod The provided code snippet includes necessary dependencies for implementing the `get_global_norm` function. Write a Python function `def get_global_norm(norm_list)` to solve the following problem: Compute total from a list of norms Here is the function: def get_global_norm(norm_list): """ Compute total from a list of norms """ total_norm = 0.0 for norm in norm_list: total_norm += norm**2.0 return sqrt(total_norm)
Compute total from a list of norms
10,215
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def is_model_parallel_parameter(p) -> bool: if hasattr(p, 'model_parallel') and p.model_parallel: return True if hasattr(p, 'tensor_model_parallel') and p.tensor_model_parallel: return True return False The provided code snippet includes necessary dependencies for implementing the `clip_grad_norm_` function. Write a Python function `def clip_grad_norm_(parameters, max_norm, norm_type=2, mpu=None)` to solve the following problem: Clips gradient norm of an iterable of parameters. This has been adapted from Nvidia megatron. We add norm averaging to consider MoE params when calculating norm as they will result in different norms across different ranks. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector). Here is the function: def clip_grad_norm_(parameters, max_norm, norm_type=2, mpu=None): """Clips gradient norm of an iterable of parameters. This has been adapted from Nvidia megatron. We add norm averaging to consider MoE params when calculating norm as they will result in different norms across different ranks. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector). """ if isinstance(parameters, torch.Tensor): parameters = [parameters] parameters = list(filter(lambda p: p.grad is not None, parameters)) max_norm = float(max_norm) norm_type = float(norm_type) if norm_type == inf: total_norm = max(p.grad.data.abs().max() for p in parameters) total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) # Take max across all GPUs. if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.MAX, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item() else: total_norm = 0 for p in parameters: if mpu is not None: if (mpu.get_model_parallel_rank() == 0) or is_model_parallel_parameter(p): param_norm = p.grad.data.norm(norm_type) total_norm += param_norm.item()**norm_type else: param_norm = p.grad.data.float().norm(norm_type) total_norm += param_norm.item()**norm_type # Sum across all model parallel GPUs. total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.SUM, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item()**(1. / norm_type) # Need to average total_norm across different GPUs due to the presence of moe params pg = groups._get_data_parallel_group() scaled_norm = total_norm * 1.0 / float(dist.get_world_size(group=pg)) scaled_norm_tensor = torch.cuda.FloatTensor([float(scaled_norm)]) dist.all_reduce(scaled_norm_tensor, group=pg) total_norm = scaled_norm_tensor.item() clip_coef = max_norm / (total_norm + 1e-6) if clip_coef < 1: for p in parameters: p.grad.data.mul_(clip_coef) return total_norm
Clips gradient norm of an iterable of parameters. This has been adapted from Nvidia megatron. We add norm averaging to consider MoE params when calculating norm as they will result in different norms across different ranks. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector).
10,216
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod if hasattr(torch.cuda, "memory_reserved"): torch_memory_reserved = torch.cuda.memory_reserved else: torch_memory_reserved = torch.cuda.memory_allocated if hasattr(torch.cuda, "max_memory_reserved"): torch_max_memory_reserved = torch.cuda.max_memory_reserved else: torch_max_memory_reserved = torch.cuda.memory_cached def is_model_parallel_parameter(p) -> bool: if hasattr(p, 'model_parallel') and p.model_parallel: return True if hasattr(p, 'tensor_model_parallel') and p.tensor_model_parallel: return True return False def bwc_tensor_model_parallel_rank(mpu=None): """Backwards-compatible way of querying the tensor model parallel rank from an ``mpu`` object. *Tensor* model parallelism means that tensors are physically split across processes. This contrasts with *pipeline* model parallelism, in which the layers are partitioned but tensors left intact. The API for tensor model parallelism has changed across versions and this helper provides a best-effort implementation across versions of ``mpu`` objects. The preferred mechanism is ``mpu.get_tensor_model_parallel_rank()``. This should "just work" with both Megatron-LM and DeepSpeed's pipeline parallelism. Args: mpu (model parallel unit, optional): The tensor model parallel rank. If ``mpu=None``, returns 0. Defaults to ``None``. Returns: int: the rank """ if mpu is None: # No model parallelism in easy :) return 0 if hasattr(mpu, 'get_tensor_model_parallel_rank'): # New Megatron and DeepSpeed convention (post pipeline-parallelism release) return mpu.get_tensor_model_parallel_rank() elif hasattr(mpu, 'get_slice_parallel_rank'): # Some DeepSpeed + pipeline parallelism versions return mpu.get_slice_parallel_rank() else: # Deprecated Megatron and DeepSpeed convention return mpu.get_model_parallel_rank() PIPE_REPLICATED = 'ds_pipe_replicated' The provided code snippet includes necessary dependencies for implementing the `get_grad_zeros` function. Write a Python function `def get_grad_zeros(parameters, mpu=None)` to solve the following problem: Compute the number of grads with zero values. This is adapted from get_grad_norm Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized Returns: Total number of params with zero values (viewed as a single vector). Here is the function: def get_grad_zeros(parameters, mpu=None): """Compute the number of grads with zero values. This is adapted from get_grad_norm Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized Returns: Total number of params with zero values (viewed as a single vector). """ if isinstance(parameters, torch.Tensor): parameters = [parameters] parameters = list(filter(lambda p: p.grad is not None, parameters)) total_zeros = 0. tensor_mp_rank = bwc_tensor_model_parallel_rank(mpu=mpu) for p in parameters: # Pipeline parallelism may replicate parameters. Avoid multi-counting. if hasattr(p, PIPE_REPLICATED) and p.ds_pipe_replicated: continue # Filter to avoid over-counting replicated tensors from tensor # model parallelism if (tensor_mp_rank > 0) and not is_model_parallel_parameter(p): continue count_zeros = p.grad.numel() - torch.count_nonzero(p.grad) total_zeros += count_zeros.item() # Sum across all model parallel GPUs. total_zeros_cuda = torch.cuda.FloatTensor([float(total_zeros)]) if mpu is not None: dist.all_reduce(total_zeros_cuda, op=dist.ReduceOp.SUM, group=mpu.get_model_parallel_group()) total_zeros = total_zeros_cuda[0].item() return total_zeros
Compute the number of grads with zero values. This is adapted from get_grad_norm Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized Returns: Total number of params with zero values (viewed as a single vector).
10,217
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod if hasattr(torch.cuda, "memory_reserved"): torch_memory_reserved = torch.cuda.memory_reserved else: torch_memory_reserved = torch.cuda.memory_allocated if hasattr(torch.cuda, "max_memory_reserved"): torch_max_memory_reserved = torch.cuda.max_memory_reserved else: torch_max_memory_reserved = torch.cuda.memory_cached def is_model_parallel_parameter(p) -> bool: if hasattr(p, 'model_parallel') and p.model_parallel: return True if hasattr(p, 'tensor_model_parallel') and p.tensor_model_parallel: return True return False def bwc_tensor_model_parallel_rank(mpu=None): """Backwards-compatible way of querying the tensor model parallel rank from an ``mpu`` object. *Tensor* model parallelism means that tensors are physically split across processes. This contrasts with *pipeline* model parallelism, in which the layers are partitioned but tensors left intact. The API for tensor model parallelism has changed across versions and this helper provides a best-effort implementation across versions of ``mpu`` objects. The preferred mechanism is ``mpu.get_tensor_model_parallel_rank()``. This should "just work" with both Megatron-LM and DeepSpeed's pipeline parallelism. Args: mpu (model parallel unit, optional): The tensor model parallel rank. If ``mpu=None``, returns 0. Defaults to ``None``. Returns: int: the rank """ if mpu is None: # No model parallelism in easy :) return 0 if hasattr(mpu, 'get_tensor_model_parallel_rank'): # New Megatron and DeepSpeed convention (post pipeline-parallelism release) return mpu.get_tensor_model_parallel_rank() elif hasattr(mpu, 'get_slice_parallel_rank'): # Some DeepSpeed + pipeline parallelism versions return mpu.get_slice_parallel_rank() else: # Deprecated Megatron and DeepSpeed convention return mpu.get_model_parallel_rank() PIPE_REPLICATED = 'ds_pipe_replicated' The provided code snippet includes necessary dependencies for implementing the `get_weight_norm` function. Write a Python function `def get_weight_norm(parameters, norm_type=2, mpu=None)` to solve the following problem: Get norm of an iterable of parameters. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Taken from Nvidia Megatron. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector). Here is the function: def get_weight_norm(parameters, norm_type=2, mpu=None): """Get norm of an iterable of parameters. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Taken from Nvidia Megatron. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector). """ if isinstance(parameters, torch.Tensor): parameters = [parameters] norm_type = float(norm_type) if norm_type == inf: total_norm = max(p.data.abs().max() for p in parameters) total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) # Take max across all GPUs. if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.MAX, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item() else: total_norm = 0. tensor_mp_rank = bwc_tensor_model_parallel_rank(mpu=mpu) for p in parameters: # Pipeline parallelism may replicate parameters. Avoid multi-counting. if hasattr(p, PIPE_REPLICATED) and p.ds_pipe_replicated: continue # Filter to avoid over-counting replicated tensors from tensor # model parallelism if (tensor_mp_rank > 0) and not is_model_parallel_parameter(p): continue param_norm = p.data.float().norm(norm_type) total_norm += param_norm**norm_type # Sum across all model parallel GPUs. total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.SUM, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item()**(1. / norm_type) if total_norm == float( 'inf') or total_norm == -float('inf') or total_norm != total_norm: total_norm = -1 return total_norm
Get norm of an iterable of parameters. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Taken from Nvidia Megatron. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector).
10,218
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def prefix_sum_inc(weights): def partition_uniform(num_items, num_parts): def _lprobe(weights, num_parts, bottleneck): def _rb_partition_balanced(weights, num_parts, eps): def partition_balanced(weights, num_parts, eps=1e-3): num_items = len(weights) # First check for the trivial edge case if num_items <= num_parts: return partition_uniform(num_items, num_parts) weights_ = prefix_sum_inc(weights) # Find the smallest bottleneck (weight of heaviest partition) bottleneck = _rb_partition_balanced(weights_, num_parts, eps=eps) # Now compute that partitioning parts, success = _lprobe(weights_, num_parts, bottleneck) assert success return parts
null
10,219
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod mem_alloced = 0 mem_cached = 0 def memory_status(msg, print_rank=-1, reset_max=False): global mem_alloced, mem_cached rank = dist.get_rank() if print_rank != -1 and rank != print_rank: return torch.cuda.synchronize() if reset_max: torch.cuda.reset_max_memory_cached() torch.cuda.reset_max_memory_allocated() new_alloced = torch.cuda.memory_allocated() new_cached = torch.cuda.memory_cached() delta_alloced = new_alloced - mem_alloced delta_cached = new_cached - mem_cached mem_cached = new_cached mem_alloced = new_alloced max_alloced = torch.cuda.max_memory_allocated() max_cached = torch.cuda.max_memory_cached() # convert to GB for printing new_alloced /= 1024**3 new_cached /= 1024**3 delta_alloced /= 1024**3 delta_cached /= 1024**3 max_alloced /= 1024**3 max_cached /= 1024**3 print( f'RANK={rank} MEMSTATS', msg, f'device={torch.cuda.current_device()} ' f'current alloc={new_alloced:0.4f}GB (delta={delta_alloced:0.4f}GB max={max_alloced:0.4f}GB) ' f'current cache={new_cached:0.4f}GB (delta={delta_cached:0.4f}GB max={max_cached:0.4f}GB)' )
null
10,220
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def get_ma_status(): if dist.is_initialized() and not dist.get_rank() == 0: return 0 return torch.cuda.memory_allocated()
null
10,221
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod if hasattr(torch.cuda, "memory_reserved"): torch_memory_reserved = torch.cuda.memory_reserved else: torch_memory_reserved = torch.cuda.memory_allocated if hasattr(torch.cuda, "max_memory_reserved"): torch_max_memory_reserved = torch.cuda.max_memory_reserved else: torch_max_memory_reserved = torch.cuda.memory_cached def see_memory_usage(message, force=False): if not force: return if dist.is_initialized() and not dist.get_rank() == 0: return # python doesn't do real-time garbage collection so do it explicitly to get the correct RAM reports gc.collect() # Print message except when distributed but not rank 0 logger.info(message) logger.info( f"MA {round(torch.cuda.memory_allocated() / (1024 * 1024 * 1024),2 )} GB \ Max_MA {round(torch.cuda.max_memory_allocated() / (1024 * 1024 * 1024),2)} GB \ CA {round(torch_memory_reserved() / (1024 * 1024 * 1024),2)} GB \ Max_CA {round(torch_max_memory_reserved() / (1024 * 1024 * 1024))} GB ") vm_stats = psutil.virtual_memory() used_GB = round(((vm_stats.total - vm_stats.available) / (1024**3)), 2) logger.info( f'CPU Virtual Memory: used = {used_GB} GB, percent = {vm_stats.percent}%') # get the peak memory to report correct data, so reset the counter for the next call if hasattr(torch.cuda, "reset_peak_memory_stats"): # pytorch 1.4+ torch.cuda.reset_peak_memory_stats()
null
10,222
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod The provided code snippet includes necessary dependencies for implementing the `call_to_str` function. Write a Python function `def call_to_str(base, *args, **kwargs)` to solve the following problem: Construct a string representation of a call. Args: base (str): name of the call args (tuple, optional): args to ``base`` kwargs (dict, optional): kwargs supplied to ``base`` Returns: str: A string representation of base(*args, **kwargs) Here is the function: def call_to_str(base, *args, **kwargs): """Construct a string representation of a call. Args: base (str): name of the call args (tuple, optional): args to ``base`` kwargs (dict, optional): kwargs supplied to ``base`` Returns: str: A string representation of base(*args, **kwargs) """ name = f'{base}(' if args: name += ', '.join(repr(arg) for arg in args) if kwargs: name += ', ' if kwargs: name += ', '.join(f'{key}={repr(arg)}' for key, arg in kwargs.items()) name += ')' return name
Construct a string representation of a call. Args: base (str): name of the call args (tuple, optional): args to ``base`` kwargs (dict, optional): kwargs supplied to ``base`` Returns: str: A string representation of base(*args, **kwargs)
10,223
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def get_only_unique_item(items): item_set = set(items) if len(item_set) != 1: raise RuntimeError(f"expected there to be only one unique element in {items}") unique_item, = item_set return unique_item
null
10,224
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def get_grad_norm(parameters, norm_type=2, mpu=None): """Get grad norm of an iterable of parameters. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Taken from Nvidia Megatron. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector). """ if isinstance(parameters, torch.Tensor): parameters = [parameters] parameters = list(filter(lambda p: p.grad is not None, parameters)) norm_type = float(norm_type) if norm_type == inf: total_norm = max(p.grad.data.abs().max() for p in parameters) total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) # Take max across all GPUs. if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.MAX, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item() else: total_norm = 0. tensor_mp_rank = bwc_tensor_model_parallel_rank(mpu=mpu) for p in parameters: # Pipeline parallelism may replicate parameters. Avoid multi-counting. if hasattr(p, PIPE_REPLICATED) and p.ds_pipe_replicated: continue # Filter to avoid over-counting replicated tensors from tensor # model parallelism if (tensor_mp_rank > 0) and not is_model_parallel_parameter(p): continue param_norm = p.grad.data.float().norm(norm_type) total_norm += param_norm.item()**norm_type # Sum across all model parallel GPUs. total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.SUM, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item()**(1. / norm_type) if total_norm == float( 'inf') or total_norm == -float('inf') or total_norm != total_norm: total_norm = -1 return total_norm The provided code snippet includes necessary dependencies for implementing the `clip_gradients` function. Write a Python function `def clip_gradients(parameters, max_norm=1.0, global_grad_norm=None, mpu=None, eps=1e-6)` to solve the following problem: Clip the gradient of a list of parameters. Args: parameters: List of parameters whose .grad will be clipped. global_grad_norm (float, optional): Precomputed gradient norm. Defaults to None. mpu (optional): model parallelism unit. Defaults to None. eps (float, optional): epsilon value added to grad norm. Defaults to 1e-6 Returns: float: the global gradient norm Here is the function: def clip_gradients(parameters, max_norm=1.0, global_grad_norm=None, mpu=None, eps=1e-6): """Clip the gradient of a list of parameters. Args: parameters: List of parameters whose .grad will be clipped. global_grad_norm (float, optional): Precomputed gradient norm. Defaults to None. mpu (optional): model parallelism unit. Defaults to None. eps (float, optional): epsilon value added to grad norm. Defaults to 1e-6 Returns: float: the global gradient norm """ if global_grad_norm is None: global_grad_norm = get_grad_norm(parameters, mpu=mpu) clip_coef = max_norm / (global_grad_norm + eps) if clip_coef < 1: for p in parameters: p.grad.detach().mul_(clip_coef) return global_grad_norm
Clip the gradient of a list of parameters. Args: parameters: List of parameters whose .grad will be clipped. global_grad_norm (float, optional): Precomputed gradient norm. Defaults to None. mpu (optional): model parallelism unit. Defaults to None. eps (float, optional): epsilon value added to grad norm. Defaults to 1e-6 Returns: float: the global gradient norm
10,225
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def get_global_norm_of_tensors(input_tensors, norm_type=2, mpu=None): """Get norm of an iterable of tensors. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Taken from Nvidia Megatron. Arguments: input_tensors (Iterable[Tensor]): an iterable of Tensors will have norm computed norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the tensors (viewed as a single vector). """ assert isinstance(input_tensors, Iterable), f'expected Iterable type not {type(input_tensors)}' assert all([torch.is_tensor(t) for t in input_tensors]), f'expected list of only tensors' norm_type = float(norm_type) if norm_type == inf: total_norm = max(t.data.abs().max() for t in input_tensors) total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.MAX, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item() else: total_norm = sum( [t.data.float().norm(norm_type).item()**norm_type for t in input_tensors]) total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) if mpu is not None: dist.all_reduce(total_norm_cuda, op=dist.ReduceOp.SUM, group=mpu.get_model_parallel_group()) total_norm = total_norm_cuda[0].item()**(1. / norm_type) if total_norm == float( 'inf') or total_norm == -float('inf') or total_norm != total_norm: total_norm = -1 return total_norm The provided code snippet includes necessary dependencies for implementing the `clip_tensors_by_global_norm` function. Write a Python function `def clip_tensors_by_global_norm(input_tensors, max_norm=1.0, global_norm=None, mpu=None, eps=1e-6)` to solve the following problem: Clip list of tensors by global norm. Args: input_tensors: List of tensors to be clipped global_norm (float, optional): Precomputed norm. Defaults to None. mpu (optional): model parallelism unit. Defaults to None. eps (float, optional): epsilon value added to grad norm. Defaults to 1e-6 Returns: float: the global norm Here is the function: def clip_tensors_by_global_norm(input_tensors, max_norm=1.0, global_norm=None, mpu=None, eps=1e-6): """Clip list of tensors by global norm. Args: input_tensors: List of tensors to be clipped global_norm (float, optional): Precomputed norm. Defaults to None. mpu (optional): model parallelism unit. Defaults to None. eps (float, optional): epsilon value added to grad norm. Defaults to 1e-6 Returns: float: the global norm """ if global_norm is None: global_norm = get_global_norm_of_tensors(input_tensors, mpu=mpu) clip_coef = max_norm / (global_norm + eps) if clip_coef < 1: for t in input_tensors: t.detach().mul_(clip_coef) return global_norm
Clip list of tensors by global norm. Args: input_tensors: List of tensors to be clipped global_norm (float, optional): Precomputed norm. Defaults to None. mpu (optional): model parallelism unit. Defaults to None. eps (float, optional): epsilon value added to grad norm. Defaults to 1e-6 Returns: float: the global norm
10,226
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def align_dense_tensors(tensor_list, alignment): num_elements = sum(t.numel() for t in tensor_list) remaining = num_elements % alignment if remaining: elements_to_add = alignment - remaining pad_tensor = torch.zeros(elements_to_add, device=tensor_list[0].device, dtype=tensor_list[0].dtype) padded_tensor_list = tensor_list + [pad_tensor] else: padded_tensor_list = tensor_list return padded_tensor_list
null
10,227
from collections.abc import Iterable from deepspeed.moe.utils import is_moe_param import os import psutil import gc from math import sqrt from math import floor from bisect import bisect_left import torch from torch._six import inf from deepspeed import comm as dist from deepspeed.utils import groups, logger from deepspeed.runtime.constants import PIPE_REPLICATED from numpy import prod def all_gather_dp_groups(partitioned_param_groups, dp_process_group, start_alignment_factor, allgather_bucket_size): for group_id, partitioned_params in enumerate(partitioned_param_groups): # Sequential AllGather Best of both worlds partition_id = dist.get_rank(group=dp_process_group[group_id]) dp_world_size = dist.get_world_size(group=dp_process_group[group_id]) num_shards = max( 1, partitioned_params[partition_id].numel() * dp_world_size // allgather_bucket_size) shard_size = partitioned_params[partition_id].numel() // num_shards # Enforce nccl/rccl alignment of start location of each shard shard_size = shard_size - (shard_size % start_alignment_factor) num_elements = shard_size assert shard_size * num_shards <= partitioned_params[partition_id].numel() for shard_id in range(num_shards): if shard_id == (num_shards - 1): num_elements = partitioned_params[partition_id].numel( ) - shard_id * shard_size shard_list = [] for dp_id in range(dp_world_size): curr_shard = partitioned_params[dp_id].narrow(0, shard_id * shard_size, num_elements).detach() shard_list.append(curr_shard) dist.all_gather(shard_list, shard_list[partition_id], dp_process_group[group_id])
null
10,228
import argparse from torch.optim import Optimizer import math from deepspeed.utils import logger def add_tuning_arguments(parser): group = parser.add_argument_group('Convergence Tuning', 'Convergence tuning configurations') # LR scheduler group.add_argument('--lr_schedule', type=str, default=None, help='LR schedule for training.') # Learning rate range test group.add_argument("--lr_range_test_min_lr", type=float, default=0.001, help='Starting lr value.') group.add_argument("--lr_range_test_step_rate", type=float, default=1.0, help='scaling rate for LR range test.') group.add_argument("--lr_range_test_step_size", type=int, default=1000, help='training steps per LR change.') group.add_argument("--lr_range_test_staircase", type=bool, default=False, help='use staircase scaling for LR range test.') # OneCycle schedule group.add_argument("--cycle_first_step_size", type=int, default=1000, help='size of first step of 1Cycle schedule (training steps).') group.add_argument("--cycle_first_stair_count", type=int, default=-1, help='first stair count for 1Cycle schedule.') group.add_argument( "--cycle_second_step_size", type=int, default=-1, help='size of second step of 1Cycle schedule (default first_step_size).') group.add_argument("--cycle_second_stair_count", type=int, default=-1, help='second stair count for 1Cycle schedule.') group.add_argument( "--decay_step_size", type=int, default=1000, help='size of intervals for applying post cycle decay (training steps).') # 1Cycle LR group.add_argument("--cycle_min_lr", type=float, default=0.01, help='1Cycle LR lower bound.') group.add_argument("--cycle_max_lr", type=float, default=0.1, help='1Cycle LR upper bound.') group.add_argument("--decay_lr_rate", type=float, default=0.0, help='post cycle LR decay rate.') # 1Cycle Momentum group.add_argument('--cycle_momentum', default=False, action='store_true', help='Enable 1Cycle momentum schedule.') group.add_argument("--cycle_min_mom", type=float, default=0.8, help='1Cycle momentum lower bound.') group.add_argument("--cycle_max_mom", type=float, default=0.9, help='1Cycle momentum upper bound.') group.add_argument("--decay_mom_rate", type=float, default=0.0, help='post cycle momentum decay rate.') # Warmup LR group.add_argument('--warmup_min_lr', type=float, default=0, help='WarmupLR minimum/initial LR value') group.add_argument('--warmup_max_lr', type=float, default=0.001, help='WarmupLR maximum LR value.') group.add_argument('--warmup_num_steps', type=int, default=1000, help='WarmupLR step count for LR warmup.') group.add_argument('--warmup_type', type=str, default=WARMUP_LOG_RATE, help='WarmupLR increasing function during warmup') return parser def parse_arguments(): parser = argparse.ArgumentParser() parser = add_tuning_arguments(parser) lr_sched_args, unknown_args = parser.parse_known_args() return lr_sched_args, unknown_args
null
10,229
import argparse from torch.optim import Optimizer import math from deepspeed.utils import logger def override_lr_range_test_params(args, params): if hasattr(args, LR_RANGE_TEST_MIN_LR) and args.lr_range_test_min_lr is not None: params[LR_RANGE_TEST_MIN_LR] = args.lr_range_test_min_lr if hasattr(args, LR_RANGE_TEST_STEP_RATE) and args.lr_range_test_step_rate is not None: params[LR_RANGE_TEST_STEP_RATE] = args.lr_range_test_step_rate if hasattr(args, LR_RANGE_TEST_STEP_SIZE) and args.lr_range_test_step_size is not None: params[LR_RANGE_TEST_STEP_SIZE] = args.lr_range_test_step_size if hasattr(args, LR_RANGE_TEST_STAIRCASE) and args.lr_range_test_staircase is not None: params[LR_RANGE_TEST_STAIRCASE] = args.lr_range_test_staircase def override_1cycle_params(args, params): if hasattr(args, CYCLE_FIRST_STEP_SIZE) and args.cycle_first_step_size is not None: params[CYCLE_FIRST_STEP_SIZE] = args.cycle_first_step_size if hasattr(args, CYCLE_FIRST_STAIR_COUNT) and args.cycle_first_stair_count is not None: params[CYCLE_FIRST_STAIR_COUNT] = args.cycle_first_stair_count if hasattr(args, CYCLE_SECOND_STEP_SIZE) and args.cycle_second_step_size is not None: params[CYCLE_SECOND_STEP_SIZE] = args.cycle_second_step_size if hasattr(args, CYCLE_SECOND_STAIR_COUNT) and args.cycle_second_stair_count is not None: params[CYCLE_SECOND_STAIR_COUNT] = args.cycle_second_stair_count if hasattr(args, DECAY_STEP_SIZE) and args.decay_step_size is not None: params[DECAY_STEP_SIZE] = args.decay_step_size # 1Cycle LR params if hasattr(args, CYCLE_MIN_LR) and args.cycle_min_lr is not None: params[CYCLE_MIN_LR] = args.cycle_min_lr if hasattr(args, CYCLE_MAX_LR) and args.cycle_max_lr is not None: params[CYCLE_MAX_LR] = args.cycle_max_lr if hasattr(args, DECAY_LR_RATE) and args.decay_lr_rate is not None: params[DECAY_LR_RATE] = args.decay_lr_rate # 1Cycle MOM params if hasattr(args, CYCLE_MIN_MOM) and args.cycle_min_mom is not None: params[CYCLE_MIN_MOM] = args.cycle_min_mom if hasattr(args, CYCLE_MAX_MOM) and args.cycle_max_mom is not None: params[CYCLE_MAX_MOM] = args.cycle_max_mom if hasattr(args, DECAY_MOM_RATE) and args.decay_mom_rate is not None: params[DECAY_MOM_RATE] = args.decay_mom_rate def override_warmupLR_params(args, params): if hasattr(args, WARMUP_MIN_LR) and args.warmup_min_lr is not None: params[WARMUP_MIN_LR] = args.warmup_min_lr if hasattr(args, WARMUP_MAX_LR) and args.warmup_max_lr is not None: params[WARMUP_MAX_LR] = args.warmup_max_lr if hasattr(args, WARMUP_NUM_STEPS) and args.warmup_num_steps is not None: params[WARMUP_NUM_STEPS] = args.warmup_num_steps if hasattr(args, WARMUP_TYPE) and args.warmup_type is not None: params[WARMUP_TYPE] = args.warmup_type def override_params(args, params): # LR range test params override_lr_range_test_params(args, params) # 1Cycle params override_1cycle_params(args, params) # WarmupLR params override_warmupLR_params(args, params)
null
10,230
import argparse from torch.optim import Optimizer import math from deepspeed.utils import logger LR_SCHEDULE = 'lr_schedule' LR_RANGE_TEST = 'LRRangeTest' ONE_CYCLE = 'OneCycle' VALID_LR_SCHEDULES = [LR_RANGE_TEST, ONE_CYCLE, WARMUP_LR, WARMUP_DECAY_LR] def override_lr_range_test_params(args, params): if hasattr(args, LR_RANGE_TEST_MIN_LR) and args.lr_range_test_min_lr is not None: params[LR_RANGE_TEST_MIN_LR] = args.lr_range_test_min_lr if hasattr(args, LR_RANGE_TEST_STEP_RATE) and args.lr_range_test_step_rate is not None: params[LR_RANGE_TEST_STEP_RATE] = args.lr_range_test_step_rate if hasattr(args, LR_RANGE_TEST_STEP_SIZE) and args.lr_range_test_step_size is not None: params[LR_RANGE_TEST_STEP_SIZE] = args.lr_range_test_step_size if hasattr(args, LR_RANGE_TEST_STAIRCASE) and args.lr_range_test_staircase is not None: params[LR_RANGE_TEST_STAIRCASE] = args.lr_range_test_staircase def override_1cycle_params(args, params): if hasattr(args, CYCLE_FIRST_STEP_SIZE) and args.cycle_first_step_size is not None: params[CYCLE_FIRST_STEP_SIZE] = args.cycle_first_step_size if hasattr(args, CYCLE_FIRST_STAIR_COUNT) and args.cycle_first_stair_count is not None: params[CYCLE_FIRST_STAIR_COUNT] = args.cycle_first_stair_count if hasattr(args, CYCLE_SECOND_STEP_SIZE) and args.cycle_second_step_size is not None: params[CYCLE_SECOND_STEP_SIZE] = args.cycle_second_step_size if hasattr(args, CYCLE_SECOND_STAIR_COUNT) and args.cycle_second_stair_count is not None: params[CYCLE_SECOND_STAIR_COUNT] = args.cycle_second_stair_count if hasattr(args, DECAY_STEP_SIZE) and args.decay_step_size is not None: params[DECAY_STEP_SIZE] = args.decay_step_size # 1Cycle LR params if hasattr(args, CYCLE_MIN_LR) and args.cycle_min_lr is not None: params[CYCLE_MIN_LR] = args.cycle_min_lr if hasattr(args, CYCLE_MAX_LR) and args.cycle_max_lr is not None: params[CYCLE_MAX_LR] = args.cycle_max_lr if hasattr(args, DECAY_LR_RATE) and args.decay_lr_rate is not None: params[DECAY_LR_RATE] = args.decay_lr_rate # 1Cycle MOM params if hasattr(args, CYCLE_MIN_MOM) and args.cycle_min_mom is not None: params[CYCLE_MIN_MOM] = args.cycle_min_mom if hasattr(args, CYCLE_MAX_MOM) and args.cycle_max_mom is not None: params[CYCLE_MAX_MOM] = args.cycle_max_mom if hasattr(args, DECAY_MOM_RATE) and args.decay_mom_rate is not None: params[DECAY_MOM_RATE] = args.decay_mom_rate def override_warmupLR_params(args, params): if hasattr(args, WARMUP_MIN_LR) and args.warmup_min_lr is not None: params[WARMUP_MIN_LR] = args.warmup_min_lr if hasattr(args, WARMUP_MAX_LR) and args.warmup_max_lr is not None: params[WARMUP_MAX_LR] = args.warmup_max_lr if hasattr(args, WARMUP_NUM_STEPS) and args.warmup_num_steps is not None: params[WARMUP_NUM_STEPS] = args.warmup_num_steps if hasattr(args, WARMUP_TYPE) and args.warmup_type is not None: params[WARMUP_TYPE] = args.warmup_type def get_config_from_args(args): if not hasattr(args, LR_SCHEDULE) or args.lr_schedule is None: return None, '--{} not specified on command line'.format(LR_SCHEDULE) if not args.lr_schedule in VALID_LR_SCHEDULES: return None, '{} is not supported LR schedule'.format(args.lr_schedule) config = {} config['type'] = args.lr_schedule config['params'] = {} if args.lr_schedule == LR_RANGE_TEST: override_lr_range_test_params(args, config['params']) elif args.lr_schedule == ONE_CYCLE: override_1cycle_params(args, config['params']) else: override_warmupLR_params(args, config['params']) return config, None
null
10,231
import argparse from torch.optim import Optimizer import math from deepspeed.utils import logger LR_RANGE_TEST = 'LRRangeTest' ONE_CYCLE = 'OneCycle' VALID_LR_SCHEDULES = [LR_RANGE_TEST, ONE_CYCLE, WARMUP_LR, WARMUP_DECAY_LR] LR_RANGE_TEST_MIN_LR = 'lr_range_test_min_lr' CYCLE_MAX_LR = 'cycle_max_lr' WARMUP_MAX_LR = 'warmup_max_lr' def get_lr_from_config(config): if not 'type' in config: return None, 'LR schedule type not defined in config' if not 'params' in config: return None, 'LR schedule params not defined in config' lr_schedule = config['type'] lr_params = config['params'] if not lr_schedule in VALID_LR_SCHEDULES: return None, '{} is not a valid LR schedule'.format(lr_schedule) if lr_schedule == LR_RANGE_TEST: return lr_params[LR_RANGE_TEST_MIN_LR], '' if lr_schedule == ONE_CYCLE: return lr_params[CYCLE_MAX_LR], '' # Warmup LR return lr_params[WARMUP_MAX_LR], ''
null
10,232
import argparse from torch.optim import Optimizer import math from deepspeed.utils import logger def get_torch_optimizer(optimizer): if isinstance(optimizer, Optimizer): return optimizer if hasattr(optimizer, 'optimizer') and isinstance(optimizer.optimizer, Optimizer): return optimizer.optimizer raise TypeError('{} is not a subclass of torch.optim.Optimizer'.format( type(optimizer).__name__))
null
10,233
import json import collections import collections.abc from functools import reduce from pydantic import BaseModel from deepspeed.utils import logger def get_list_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value)
null
10,234
import json import collections import collections.abc from functools import reduce from pydantic import BaseModel from deepspeed.utils import logger def get_dict_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value)
null
10,235
import math from typing import List import torch from torch import Tensor from deepspeed import comm as dist from torch.distributed import ProcessGroup import torch.nn.functional from deepspeed.utils import instrument_w_nvtx def _torch_reduce_scatter_fn(input_tensor: Tensor, output_tensor: Tensor, group=None, async_op=False, prof=False): return instrument_w_nvtx(dist.reduce_scatter_fn)(output_tensor, input_tensor, group=group, async_op=async_op) The provided code snippet includes necessary dependencies for implementing the `reduce_scatter_coalesced` function. Write a Python function `def reduce_scatter_coalesced( tensors: List[Tensor], group: ProcessGroup = None, ) -> List[Tensor]` to solve the following problem: simultaneously reduce-scatter a list of tensors - this can be done more efficiently than individual reduce scatter calls TODO. see if PyTorch team wants a c++ version of this for ProcessGroupNCCL Here is the function: def reduce_scatter_coalesced( tensors: List[Tensor], group: ProcessGroup = None, ) -> List[Tensor]: """simultaneously reduce-scatter a list of tensors - this can be done more efficiently than individual reduce scatter calls TODO. see if PyTorch team wants a c++ version of this for ProcessGroupNCCL """ this_rank = dist.get_rank(group) world_sz = dist.get_world_size(group) partition_lst_for_each_tensor = [None] * len(tensors) for tensor_idx, tensor in enumerate(tensors): flattened_tensor = tensor.view(-1) chunk_sz = math.ceil(tensor.numel() / world_sz) partition_lst_for_each_tensor[tensor_idx] = [ flattened_tensor[rank * chunk_sz:rank * chunk_sz + chunk_sz] for rank in range(0, world_sz) ] padded_partition_sz_for_each_tensor = tuple( math.ceil(t.numel() / world_sz) for t in tensors) if len(tensors) == 1 and tensors[0].numel() % world_sz == 0: # if there's only one tensor being reduced and we don't need to pad # we have an opportunity to avoid a memory allocation tensor_partition_flat_buffer = tensors[0].view(-1) else: # interleave tensor partitions such that the correct reduced partitions of each tensor # end up at each rank tensor_partitions_lst_with_padding = [] for rank in range(world_sz): for tensor_idx in range(len(tensors)): # add tensor content tensor_chunk = partition_lst_for_each_tensor[tensor_idx][rank] tensor_partitions_lst_with_padding.append(tensor_chunk) # add padding if necessary padding_sz = padded_partition_sz_for_each_tensor[ tensor_idx] - tensor_chunk.numel() if padding_sz > 0: tensor_partitions_lst_with_padding.append( torch.empty(padding_sz, dtype=tensor_chunk.dtype, device=tensor_chunk.device)) tensor_partition_flat_buffer = instrument_w_nvtx( torch.cat)(tensor_partitions_lst_with_padding) tensor_partition_flat_buffer.div_(world_sz) # pre-divide tensor_partition_buffer_for_each_rank: List[Tensor] = torch.chunk( tensor_partition_flat_buffer, world_sz) # batched reduce-scatter call _torch_reduce_scatter_fn(tensor_partition_flat_buffer, tensor_partition_buffer_for_each_rank[this_rank], group=group) # reverse procedure of the interleaving done previously, done on the # result of the batched reduce-scatter output_lst: List[Tensor] = [None] * len(tensors) offset = 0 for tensor_idx in range(len(tensors)): output_lst[tensor_idx] = tensor_partition_buffer_for_each_rank[this_rank].narrow( 0, offset, partition_lst_for_each_tensor[tensor_idx][this_rank].numel()) offset += padded_partition_sz_for_each_tensor[tensor_idx] return output_lst
simultaneously reduce-scatter a list of tensors - this can be done more efficiently than individual reduce scatter calls TODO. see if PyTorch team wants a c++ version of this for ProcessGroupNCCL
10,236
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config CURRICULUM_LEARNING = "curriculum_learning" CURRICULUM_ENABLED = "enabled" CURRICULUM_ENABLED_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): def get_curriculum_enabled(param_dict): if CURRICULUM_LEARNING in param_dict.keys(): return get_scalar_param(param_dict[CURRICULUM_LEARNING], CURRICULUM_ENABLED, CURRICULUM_ENABLED_DEFAULT) else: return False
null
10,237
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config CURRICULUM_LEARNING = "curriculum_learning" CURRICULUM_ENABLED = "enabled" def get_curriculum_params(param_dict): if CURRICULUM_LEARNING in param_dict.keys(): curriculum_params = copy.copy(param_dict[CURRICULUM_LEARNING]) curriculum_params.pop(CURRICULUM_ENABLED) return curriculum_params else: return False
null
10,238
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config PROGRESSIVE_LAYER_DROP = "progressive_layer_drop" PLD_ENABLED = "enabled" PLD_ENABLED_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_pld_enabled(param_dict): if PROGRESSIVE_LAYER_DROP in param_dict.keys(): return get_scalar_param(param_dict[PROGRESSIVE_LAYER_DROP], PLD_ENABLED, PLD_ENABLED_DEFAULT) else: return False
null
10,239
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config PROGRESSIVE_LAYER_DROP = "progressive_layer_drop" PLD_ENABLED = "enabled" def get_pld_params(param_dict): if PROGRESSIVE_LAYER_DROP in param_dict.keys(): pld_params = copy.copy(param_dict[PROGRESSIVE_LAYER_DROP]) pld_params.pop(PLD_ENABLED) return pld_params else: return False
null
10,240
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config AMP = "amp" AMP_ENABLED = "enabled" AMP_ENABLED_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_amp_enabled(param_dict): if AMP in param_dict.keys(): return get_scalar_param(param_dict[AMP], AMP_ENABLED, AMP_ENABLED_DEFAULT) else: return False
null
10,241
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config AMP = "amp" AMP_ENABLED = "enabled" def get_amp_params(param_dict): if AMP in param_dict.keys(): amp_params = copy.copy(param_dict[AMP]) amp_params.pop(AMP_ENABLED) return amp_params else: return False
null
10,242
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config def get_fp16_enabled(param_dict): if FP16 in param_dict.keys(): return get_scalar_param(param_dict[FP16], FP16_ENABLED, FP16_ENABLED_DEFAULT) else: return False FP16 = "fp16" FP16_MASTER_WEIGHTS_AND_GRADS = "fp16_master_weights_and_grads" FP16_MASTER_WEIGHTS_AND_GRADS_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_fp16_master_weights_and_grads_enabled(param_dict): if get_fp16_enabled(param_dict): return get_scalar_param(param_dict[FP16], FP16_MASTER_WEIGHTS_AND_GRADS, FP16_MASTER_WEIGHTS_AND_GRADS_DEFAULT) else: return False
null
10,243
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config def get_fp16_enabled(param_dict): FP16 = "fp16" FP16_AUTO_CAST = "auto_cast" FP16_AUTO_CAST_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): def get_fp16_auto_cast(param_dict): if get_fp16_enabled(param_dict): return get_scalar_param(param_dict[FP16], FP16_AUTO_CAST, FP16_AUTO_CAST_DEFAULT)
null
10,244
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config def get_fp16_enabled(param_dict): if FP16 in param_dict.keys(): return get_scalar_param(param_dict[FP16], FP16_ENABLED, FP16_ENABLED_DEFAULT) else: return False def get_bfloat16_enabled(param_dict): for key in [BFLOAT16, BFLOAT16_OLD]: if key in param_dict.keys(): return get_scalar_param(param_dict[key], BFLOAT16_ENABLED, BFLOAT16_ENABLED_DEFAULT) return False FP16 = "fp16" FP16_LOSS_SCALE = "loss_scale" FP16_LOSS_SCALE_DEFAULT = 0 def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_loss_scale(param_dict): if get_fp16_enabled(param_dict): return get_scalar_param(param_dict[FP16], FP16_LOSS_SCALE, FP16_LOSS_SCALE_DEFAULT) elif get_bfloat16_enabled(param_dict): return 1.0 else: return FP16_LOSS_SCALE_DEFAULT
null
10,245
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config def get_fp16_enabled(param_dict): if FP16 in param_dict.keys(): return get_scalar_param(param_dict[FP16], FP16_ENABLED, FP16_ENABLED_DEFAULT) else: return False def get_bfloat16_enabled(param_dict): for key in [BFLOAT16, BFLOAT16_OLD]: if key in param_dict.keys(): return get_scalar_param(param_dict[key], BFLOAT16_ENABLED, BFLOAT16_ENABLED_DEFAULT) return False FP16 = "fp16" FP16_INITIAL_SCALE_POWER = "initial_scale_power" FP16_INITIAL_SCALE_POWER_DEFAULT = 32 def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_initial_dynamic_scale(param_dict): if get_fp16_enabled(param_dict): initial_scale_power = get_scalar_param(param_dict[FP16], FP16_INITIAL_SCALE_POWER, FP16_INITIAL_SCALE_POWER_DEFAULT) elif get_bfloat16_enabled(param_dict): initial_scale_power = 0 else: initial_scale_power = FP16_INITIAL_SCALE_POWER_DEFAULT return 2**initial_scale_power
null
10,246
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config def get_fp16_enabled(param_dict): if FP16 in param_dict.keys(): return get_scalar_param(param_dict[FP16], FP16_ENABLED, FP16_ENABLED_DEFAULT) else: return False FP16 = "fp16" FP16_INITIAL_SCALE_POWER = "initial_scale_power" FP16_INITIAL_SCALE_POWER_DEFAULT = 32 FP16_LOSS_SCALE_WINDOW = "loss_scale_window" FP16_LOSS_SCALE_WINDOW_DEFAULT = 1000 FP16_HYSTERESIS = "hysteresis" FP16_HYSTERESIS_DEFAULT = 2 FP16_MIN_LOSS_SCALE = "min_loss_scale" FP16_MIN_LOSS_SCALE_DEFAULT = 1 INITIAL_LOSS_SCALE = 'init_scale' SCALE_WINDOW = 'scale_window' DELAYED_SHIFT = 'delayed_shift' MIN_LOSS_SCALE = 'min_scale' def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_dynamic_loss_scale_args(param_dict): loss_scale_args = None if get_fp16_enabled(param_dict): fp16_dict = param_dict[FP16] dynamic_loss_args = [ FP16_INITIAL_SCALE_POWER, FP16_LOSS_SCALE_WINDOW, FP16_MIN_LOSS_SCALE, FP16_HYSTERESIS, ] if any(arg in list(fp16_dict.keys()) for arg in dynamic_loss_args): init_scale = get_scalar_param(fp16_dict, FP16_INITIAL_SCALE_POWER, FP16_INITIAL_SCALE_POWER_DEFAULT) scale_window = get_scalar_param(fp16_dict, FP16_LOSS_SCALE_WINDOW, FP16_LOSS_SCALE_WINDOW_DEFAULT) delayed_shift = get_scalar_param(fp16_dict, FP16_HYSTERESIS, FP16_HYSTERESIS_DEFAULT) min_loss_scale = get_scalar_param(fp16_dict, FP16_MIN_LOSS_SCALE, FP16_MIN_LOSS_SCALE_DEFAULT) loss_scale_args = { INITIAL_LOSS_SCALE: 2**init_scale, SCALE_WINDOW: scale_window, DELAYED_SHIFT: delayed_shift, MIN_LOSS_SCALE: min_loss_scale, } return loss_scale_args
null
10,247
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config GRADIENT_ACCUMULATION_STEPS = "gradient_accumulation_steps" GRADIENT_ACCUMULATION_STEPS_DEFAULT = None def get_scalar_param(param_dict, param_name, param_default_value): def get_gradient_accumulation_steps(param_dict): return get_scalar_param(param_dict, GRADIENT_ACCUMULATION_STEPS, GRADIENT_ACCUMULATION_STEPS_DEFAULT)
null
10,248
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config SPARSE_GRADIENTS = "sparse_gradients" SPARSE_GRADIENTS_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_sparse_gradients_enabled(param_dict): return get_scalar_param(param_dict, SPARSE_GRADIENTS, SPARSE_GRADIENTS_DEFAULT)
null
10,249
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config COMMUNICATION_DATA_TYPE = "communication_data_type" COMMUNICATION_DATA_TYPE_DEFAULT = None def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_communication_data_type(param_dict): val = get_scalar_param(param_dict, COMMUNICATION_DATA_TYPE, COMMUNICATION_DATA_TYPE_DEFAULT) val = val.lower() if val is not None else val if val is None: return val # we must determine it by other parameters elif val == "fp32": return torch.float32 elif val == "fp16": return torch.float16 elif val == "bfp16": return torch.bfloat16 raise ValueError( f"Invalid communication_data_type. Supported data types: ['fp16', 'bfp16', 'fp32']. Got: {val}" )
null
10,250
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config PRESCALE_GRADIENTS = "prescale_gradients" PRESCALE_GRADIENTS_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): def get_prescale_gradients(param_dict): return get_scalar_param(param_dict, PRESCALE_GRADIENTS, PRESCALE_GRADIENTS_DEFAULT)
null
10,251
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config GRADIENT_PREDIVIDE_FACTOR = "gradient_predivide_factor" GRADIENT_PREDIVIDE_FACTOR_DEFAULT = 1.0 def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_gradient_predivide_factor(param_dict): return get_scalar_param(param_dict, GRADIENT_PREDIVIDE_FACTOR, GRADIENT_PREDIVIDE_FACTOR_DEFAULT)
null
10,252
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config STEPS_PER_PRINT = "steps_per_print" STEPS_PER_PRINT_DEFAULT = 10 def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_steps_per_print(param_dict): return get_scalar_param(param_dict, STEPS_PER_PRINT, STEPS_PER_PRINT_DEFAULT)
null
10,253
import os from typing import Union import torch import json import copy import base64 from .constants import * from .fp16.loss_scaler import ( INITIAL_LOSS_SCALE, SCALE_WINDOW, DELAYED_SHIFT, MIN_LOSS_SCALE, ) from .config_utils import ( get_scalar_param, dict_raise_error_on_duplicate_keys, ScientificNotationEncoder, ) from .zero.config import get_zero_config, ZeroStageEnum from .activation_checkpointing.config import DeepSpeedActivationCheckpointingConfig from ..comm.config import DeepSpeedCommsConfig from ..monitor.config import DeepSpeedMonitorConfig from deepspeed import comm as dist from ..git_version_info import version as __version__ from ..utils import logger from ..elasticity import ( elasticity_enabled, compute_elastic_config, ensure_immutable_elastic_config, ) from ..elasticity.config import ElasticityConfigError from ..elasticity.constants import ( ELASTICITY, IGNORE_NON_ELASTIC_BATCH_INFO, IGNORE_NON_ELASTIC_BATCH_INFO_DEFAULT, MODEL_PARLLEL_SIZE, MODEL_PARLLEL_SIZE_DEFAULT, NUM_GPUS_PER_NODE, NUM_GPUS_PER_NODE_DEFAULT, ) from ..profiling.config import DeepSpeedFlopsProfilerConfig from ..autotuning.config import DeepSpeedAutotuningConfig from ..nebula.config import DeepSpeedNebulaConfig from ..compression.config import get_compression_config, get_quantize_enabled from ..compression.constants import * from .swap_tensor.aio_config import get_aio_config DISABLE_ALLGATHER = "disable_allgather" DISABLE_ALLGATHER_DEFAULT = False def get_scalar_param(param_dict, param_name, param_default_value): return param_dict.get(param_name, param_default_value) def get_disable_allgather(param_dict): return get_scalar_param(param_dict, DISABLE_ALLGATHER, DISABLE_ALLGATHER_DEFAULT)
null