Barbados / bundle /source /multipack_sampler.py

Release Barbados looped MoE research prototype

d5147ba verified 15 days ago

5.61 kB

from typing import Optional

import torch.distributed as dist
from torch.utils.data import Sampler

import numpy as np
import numba


@numba.njit
def lpt_check(heap: np.ndarray, A: np.ndarray, c: int, n: int):
    # LPT (Longest processing time first scheduling)
    # Time: O(|A| log |A| + |A| log n)

    A = np.sort(A)[::-1]
    heap.fill(0)
    for size in A:
        # Put into smallest element
        heap[1] += size
        if heap[1] > c:
            return False

        # Heapify (Sink)
        # https://stackoverflow.com/questions/20397674/replacing-element-in-min-heap
        u = 1
        while (u << 1) <= n:
            v = u << 1  # lch
            rch = (u << 1) | 1
            if rch <= n and heap[rch] < heap[v]:
                v = rch
            
            if heap[u] <= heap[v]:
                break

            heap[u], heap[v] = heap[v], heap[u]
            u = v

    return True


@numba.njit
def lpt_with_result(heap: np.ndarray, A: np.ndarray, n: int, rank: int):
    # LPT (Longest processing time first scheduling)
    # Time: O(|A| log |A| + |A| log n)

    result = []

    indices = np.argsort(A, kind="mergesort")[::-1]  # Stable sort for coherence across machine and versions
    A = A[indices]

    heap.fill(0)
    heap_id = np.arange(-1, n, dtype=A.dtype)
    for idx, size in enumerate(A):
        # Put into smallest element
        heap[1] += size
        if heap_id[1] == rank:
            result.append(indices[idx])

        # Heapify (Sink)
        # https://stackoverflow.com/questions/20397674/replacing-element-in-min-heap
        u = 1
        while (u << 1) <= n:
            v = u << 1  # lch
            rch = (u << 1) | 1
            if rch <= n and heap[rch] < heap[v]:
                v = rch
            
            if heap[u] <= heap[v]:
                break

            heap[u], heap[v] = heap[v], heap[u]
            heap_id[u], heap_id[v] = heap_id[v], heap_id[u]
            u = v

    return np.array(result)


@numba.njit
def allocate(heap: np.ndarray, start_index: int, lengths: np.ndarray, rank: int, c: int, n: int):
    # Dynamic batch allocator, binary search + LPT
    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)

    # Linear scan
    total_len = 0
    end_index = start_index
    while end_index < lengths.size:
        l = lengths[end_index]
        total_len += l
        end_index += 1

        if total_len >= c * n:
            break

    a = lengths[start_index: end_index]

    # binary search [l, r)
    l = 1
    r = 1 + a.size

    while r - l > 1:
        m = (l + r) // 2
        if lpt_check(heap, a[:m], c, n):
            l = m
        else:
            r = m

    # use length l
    batch = start_index + lpt_with_result(heap, a[:l], n, rank)

    return l >= n, l, batch, np.sum(a[:l])


class MultipackDistributedBatchSampler(Sampler):
    """Unpadded length sampling using Multipack V2, for models with quadratic attention complexity.
       It also tries to evenly distribute the sequences using LPT, so that quadratic load is more balanced.
       Approximate (at most 1.33x ?) the optimal solution of the identical-machines scheduling problem, which is NP-hard.
       Time Complexity: O(n log n log k)
       n = maximum number of sequences per batch, k = number of nodes
    """

    def __init__(
        self,
        batch_max_length: int,
        lengths: np.ndarray,
        num_replicas: Optional[int] = None,
        rank: Optional[int] = None,
        drop_last_batch: bool = False
    ):
        # Get rank
        if num_replicas is None:
            if not dist.is_available():
                raise RuntimeError("Requires distributed package to be available")
            num_replicas = dist.get_world_size()
        if rank is None:
            if not dist.is_available():
                raise RuntimeError("Requires distributed package to be available")
            rank = dist.get_rank()

        self.num_replicas = num_replicas
        self.rank = rank
        self.drop_last_batch = drop_last_batch

        self.batch_max_length = batch_max_length
        self.lengths = lengths
        assert isinstance(self.lengths, np.ndarray)

        # statistics
        self.eff_total_used = 0
        self.eff_total_slots = 0

    def iter(self, start_index: int = 0):
        # Allocate workspace
        heap = np.empty(self.num_replicas + 1, dtype=self.lengths.dtype)

        # Cursor after the last yielded batch (next batch starts here). O(1) resume.
        self.cursor = int(start_index)
        while start_index < self.lengths.size:
            is_full, global_numseq, batch, batch_totlen = allocate(heap,
                                                                   start_index, self.lengths,
                                                                   rank=self.rank, c=self.batch_max_length, n=self.num_replicas)
            start_index += global_numseq
            self.cursor = int(start_index)

            if not is_full:  # Skip batch with empty NPUs
                break
            if self.drop_last_batch and (start_index >= self.lengths.size):  # Skip last batch
                break

            self.eff_total_used += batch_totlen
            self.eff_total_slots += self.num_replicas * self.batch_max_length
            
            yield batch
            
    def estimate_num_batches(self):
        # Rough estimate, don't consider packing efficiency
        return round(np.sum(self.lengths) / (self.num_replicas * self.batch_max_length))

    def efficiency(self):
        return self.eff_total_used / self.eff_total_slots

	from typing import Optional

	import torch.distributed as dist
	from torch.utils.data import Sampler

	import numpy as np
	import numba


	@numba.njit
	def lpt_check(heap: np.ndarray, A: np.ndarray, c: int, n: int):
	# LPT (Longest processing time first scheduling)
	# Time: O(\|A\| log \|A\| + \|A\| log n)

	A = np.sort(A)[::-1]
	heap.fill(0)
	for size in A:
	# Put into smallest element
	heap[1] += size
	if heap[1] > c:
	return False

	# Heapify (Sink)
	# https://stackoverflow.com/questions/20397674/replacing-element-in-min-heap
	u = 1
	while (u << 1) <= n:
	v = u << 1 # lch
	rch = (u << 1) \| 1
	if rch <= n and heap[rch] < heap[v]:
	v = rch

	if heap[u] <= heap[v]:
	break

	heap[u], heap[v] = heap[v], heap[u]
	u = v

	return True


	@numba.njit
	def lpt_with_result(heap: np.ndarray, A: np.ndarray, n: int, rank: int):
	# LPT (Longest processing time first scheduling)
	# Time: O(\|A\| log \|A\| + \|A\| log n)

	result = []

	indices = np.argsort(A, kind="mergesort")[::-1] # Stable sort for coherence across machine and versions
	A = A[indices]

	heap.fill(0)
	heap_id = np.arange(-1, n, dtype=A.dtype)
	for idx, size in enumerate(A):
	# Put into smallest element
	heap[1] += size
	if heap_id[1] == rank:
	result.append(indices[idx])

	# Heapify (Sink)
	# https://stackoverflow.com/questions/20397674/replacing-element-in-min-heap
	u = 1
	while (u << 1) <= n:
	v = u << 1 # lch
	rch = (u << 1) \| 1
	if rch <= n and heap[rch] < heap[v]:
	v = rch

	if heap[u] <= heap[v]:
	break

	heap[u], heap[v] = heap[v], heap[u]
	heap_id[u], heap_id[v] = heap_id[v], heap_id[u]
	u = v

	return np.array(result)


	@numba.njit
	def allocate(heap: np.ndarray, start_index: int, lengths: np.ndarray, rank: int, c: int, n: int):
	# Dynamic batch allocator, binary search + LPT
	# ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)

	# Linear scan
	total_len = 0
	end_index = start_index
	while end_index < lengths.size:
	l = lengths[end_index]
	total_len += l
	end_index += 1

	if total_len >= c * n:
	break

	a = lengths[start_index: end_index]

	# binary search [l, r)
	l = 1
	r = 1 + a.size

	while r - l > 1:
	m = (l + r) // 2
	if lpt_check(heap, a[:m], c, n):
	l = m
	else:
	r = m

	# use length l
	batch = start_index + lpt_with_result(heap, a[:l], n, rank)

	return l >= n, l, batch, np.sum(a[:l])


	class MultipackDistributedBatchSampler(Sampler):
	"""Unpadded length sampling using Multipack V2, for models with quadratic attention complexity.
	It also tries to evenly distribute the sequences using LPT, so that quadratic load is more balanced.

	Approximate (at most 1.33x ?) the optimal solution of the identical-machines scheduling problem, which is NP-hard.

	Time Complexity: O(n log n log k)
	n = maximum number of sequences per batch, k = number of nodes
	"""

	def __init__(
	self,
	batch_max_length: int,
	lengths: np.ndarray,
	num_replicas: Optional[int] = None,
	rank: Optional[int] = None,
	drop_last_batch: bool = False
	):
	# Get rank
	if num_replicas is None:
	if not dist.is_available():
	raise RuntimeError("Requires distributed package to be available")
	num_replicas = dist.get_world_size()
	if rank is None:
	if not dist.is_available():
	raise RuntimeError("Requires distributed package to be available")
	rank = dist.get_rank()

	self.num_replicas = num_replicas
	self.rank = rank
	self.drop_last_batch = drop_last_batch

	self.batch_max_length = batch_max_length
	self.lengths = lengths
	assert isinstance(self.lengths, np.ndarray)

	# statistics
	self.eff_total_used = 0
	self.eff_total_slots = 0

	def iter(self, start_index: int = 0):
	# Allocate workspace
	heap = np.empty(self.num_replicas + 1, dtype=self.lengths.dtype)

	# Cursor after the last yielded batch (next batch starts here). O(1) resume.
	self.cursor = int(start_index)
	while start_index < self.lengths.size:
	is_full, global_numseq, batch, batch_totlen = allocate(heap,
	start_index, self.lengths,
	rank=self.rank, c=self.batch_max_length, n=self.num_replicas)
	start_index += global_numseq
	self.cursor = int(start_index)

	if not is_full: # Skip batch with empty NPUs
	break
	if self.drop_last_batch and (start_index >= self.lengths.size): # Skip last batch
	break

	self.eff_total_used += batch_totlen
	self.eff_total_slots += self.num_replicas * self.batch_max_length

	yield batch

	def estimate_num_batches(self):
	# Rough estimate, don't consider packing efficiency
	return round(np.sum(self.lengths) / (self.num_replicas * self.batch_max_length))

	def efficiency(self):
	return self.eff_total_used / self.eff_total_slots