File size: 3,936 Bytes
0f775e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
"""Reference-EQUIVALENT submission for `dist-tp-embedding-allreduce` — validation only, NOT shipped.

Two implementations, selected by EMB_MODE ("sparse" | "dense"), so the cost of all-reducing a mostly-zero
activation can be measured with everything else held fixed:

  dense  : mask out-of-range ids to zero and exchange the whole (T, H) activation, then add   -> 2*T*H*2 link
  sparse : both ranks derive the owned-position sets from the REPLICATED ids with no communication, exchange
           only the compact blocks of owned rows, and scatter into place                      ->   T*H*2 link

Neither uses a torch.distributed data collective or a pre-packaged fused collective.  Transport is a shared
MAPPED HOST buffer (mmap of /dev/shm + cudaHostRegister PORTABLE|MAPPED) with sequence-counter flags.
"""
import mmap
import os

import torch

MAX_BYTES = 1 << 29          # 512 MiB per slot
MODE = os.environ.get("EMB_MODE", "sparse")
_S = {}


def _map(path, nbytes, ctx, register):
    if ctx.rank == 0:
        with open(path, "wb") as f:
            f.truncate(nbytes)
    ctx.barrier()
    f = open(path, "r+b")
    mm = mmap.mmap(f.fileno(), nbytes)
    t = torch.frombuffer(mm, dtype=torch.uint8)
    if register:
        err = torch.cuda.cudart().cudaHostRegister(t.data_ptr(), nbytes, 3)
        assert int(err) == 0, f"cudaHostRegister failed: {err}"
    ctx.barrier()
    if ctx.rank == 0:
        os.unlink(path)
    return f, mm, t


def _setup(ctx):
    if "data" in _S:
        return
    tag = os.environ.get("MASTER_PORT", "0")
    _S["data"] = _map(f"/dev/shm/emb_data_{tag}", ctx.world_size * MAX_BYTES, ctx, True)
    _S["flag"] = _map(f"/dev/shm/emb_flag_{tag}", 4096, ctx, False)
    _S["flags"] = _S["flag"][2].view(torch.int64)
    _S["flags"].zero_()
    _S["seq"] = 0
    ctx.barrier()


def _wait(flags, i, seq):
    while int(flags[i]) < seq:
        pass


def tp_embedding_allreduce(ids, weight, ctx):
    _setup(ctx)
    Vl, H = weight.shape
    T = ids.numel()
    W, r = ctx.world_size, ctx.rank
    peer = 1 - r
    lo = r * Vl
    flags = _S["flags"]
    _S["seq"] += 1
    seq = _S["seq"]
    raw = _S["data"][2].view(torch.bfloat16).view(W, -1)

    if MODE == "dense":
        local_ids = ids - lo
        inside = (local_ids >= 0) & (local_ids < Vl)
        rows = weight[local_ids.clamp(0, Vl - 1)]
        rows = torch.where(inside.unsqueeze(-1), rows,
                           torch.zeros((), dtype=rows.dtype, device=rows.device))
        slot = raw[:, :T * H].view(W, T, H)
        slot[r].copy_(rows, non_blocking=True)
        torch.cuda.synchronize()
        flags[r] = seq
        _wait(flags, peer, seq)
        buf = torch.empty(T, H, dtype=weight.dtype, device=weight.device)
        buf.copy_(slot[peer], non_blocking=True)
        out = rows + buf                                     # exactly one side is non-zero per row
        torch.cuda.synchronize()
    else:
        owner = torch.div(ids, Vl, rounding_mode="floor")     # replicated -> both ranks agree, no comms
        pos_mine = (owner == r).nonzero(as_tuple=True)[0]
        pos_peer = (owner == peer).nonzero(as_tuple=True)[0]
        n = pos_mine.numel()
        block = weight[ids[pos_mine] - lo]                    # (n, H) compact, only rows I own
        slot = raw[:, :T * H].view(W, T, H)
        if n:
            slot[r][:n].copy_(block, non_blocking=True)
        torch.cuda.synchronize()
        flags[r] = seq
        _wait(flags, peer, seq)
        out = torch.empty(T, H, dtype=weight.dtype, device=weight.device)
        if n:
            out[pos_mine] = block
        m = pos_peer.numel()
        if m:
            buf = torch.empty(m, H, dtype=weight.dtype, device=weight.device)
            buf.copy_(slot[peer][:m], non_blocking=True)
            out[pos_peer] = buf
        torch.cuda.synchronize()

    flags[W + r] = seq
    _wait(flags, W + peer, seq)
    return out.to(torch.bfloat16)