|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | +"""Expert-parallel (EP) slice helpers. |
| 3 | +
|
| 4 | +Under expert parallelism each rank only *uses* the routed experts it owns, yet |
| 5 | +file-granular loading makes every rank read the whole shard -- the unowned |
| 6 | +experts' bytes are read and then discarded. These helpers build a tensor-name |
| 7 | +predicate selecting just this rank's owned experts (plus every non-expert |
| 8 | +tensor), so a partial-read-capable loader can skip the unowned bytes: |
| 9 | +
|
| 10 | + from fastsafetensors import SafeTensorsFileLoader |
| 11 | + from fastsafetensors.ep_slice import expert_parallel_filter |
| 12 | +
|
| 13 | + loader = SafeTensorsFileLoader(pg, device, nogds=True) |
| 14 | + loader.set_tensor_filter(expert_parallel_filter(num_experts=256, |
| 15 | + ep_size=2, ep_rank=rank)) |
| 16 | + loader.add_filenames(...) |
| 17 | + bufs = loader.copy_files_to_device() |
| 18 | +
|
| 19 | +Owned experts use contiguous-block ("linear") assignment: each rank owns |
| 20 | +``num_experts // ep_size`` consecutive experts, with any remainder given to the |
| 21 | +lowest-numbered ranks. This is a common expert-to-rank convention; the caller is |
| 22 | +responsible for ensuring it matches the assignment its runtime expects. No |
| 23 | +external dependency is required. |
| 24 | +""" |
| 25 | +import os |
| 26 | +import re |
| 27 | +from typing import Callable, Optional, Pattern, Tuple |
| 28 | + |
| 29 | +# Matches the per-expert index in routed-MoE tensor names, e.g. |
| 30 | +# "model.layers.3.mlp.experts.42.w1.weight" or DeepSeek's |
| 31 | +# "...ffn.experts.42.gate_proj.weight". Override for a different convention. |
| 32 | +DEFAULT_EXPERT_PATTERN: Pattern[str] = re.compile(r"\.experts\.(\d+)\.") |
| 33 | + |
| 34 | + |
| 35 | +def owned_expert_range(num_experts: int, ep_size: int, ep_rank: int) -> Tuple[int, int]: |
| 36 | + """Return the ``[lo, hi)`` routed-expert indices owned by ``ep_rank``. |
| 37 | +
|
| 38 | + Contiguous-block ("linear") assignment: each rank owns a consecutive block |
| 39 | + of experts, with the remainder distributed to the lowest-numbered ranks. |
| 40 | + """ |
| 41 | + if ep_size <= 0: |
| 42 | + raise ValueError(f"ep_size must be positive, got {ep_size}") |
| 43 | + if not 0 <= ep_rank < ep_size: |
| 44 | + raise ValueError(f"ep_rank {ep_rank} out of range for ep_size {ep_size}") |
| 45 | + base = num_experts // ep_size |
| 46 | + rem = num_experts % ep_size |
| 47 | + local = base + (1 if ep_rank < rem else 0) |
| 48 | + start = ep_rank * base + min(ep_rank, rem) |
| 49 | + return (start, start + local) |
| 50 | + |
| 51 | + |
| 52 | +def expert_parallel_filter( |
| 53 | + num_experts: int, |
| 54 | + ep_size: int, |
| 55 | + ep_rank: int, |
| 56 | + pattern: Pattern[str] = DEFAULT_EXPERT_PATTERN, |
| 57 | +) -> Callable[[str], bool]: |
| 58 | + """Build a ``keep(name) -> bool`` predicate for this EP rank. |
| 59 | +
|
| 60 | + Non-expert tensors (names not matching ``pattern``) are kept on every rank; |
| 61 | + routed-expert tensors are kept only when their index is in this rank's owned |
| 62 | + range. Pass the predicate to ``SafeTensorsFileLoader.set_tensor_filter`` or |
| 63 | + ``SafeTensorsMetadata.select_byte_ranges``. |
| 64 | + """ |
| 65 | + lo, hi = owned_expert_range(num_experts, ep_size, ep_rank) |
| 66 | + |
| 67 | + def keep(name: str) -> bool: |
| 68 | + m = pattern.search(name) |
| 69 | + if m is None: |
| 70 | + return True |
| 71 | + return lo <= int(m.group(1)) < hi |
| 72 | + |
| 73 | + return keep |
| 74 | + |
| 75 | + |
| 76 | +def expert_parallel_filter_from_env() -> Optional[Callable[[str], bool]]: |
| 77 | + """Build an EP filter from environment variables, or ``None`` if disabled. |
| 78 | +
|
| 79 | + Recognized variables (kept compatible with the DGX Spark overlay this |
| 80 | + prototype generalizes): |
| 81 | +
|
| 82 | + ``FASTSAFETENSORS_EP_SLICE=1`` enable EP-slice reading |
| 83 | + ``FASTSAFETENSORS_EP_NUM_EXPERTS=N`` global routed-expert count (required) |
| 84 | + ``FASTSAFETENSORS_EP_SIZE`` / ``_RANK`` override EP size/rank; otherwise |
| 85 | + taken from the initialized |
| 86 | + torch.distributed group, else from |
| 87 | + ``WORLD_SIZE`` / ``RANK``. |
| 88 | +
|
| 89 | + Returns ``None`` (load everything) unless EP-slice is enabled, the expert |
| 90 | + count is known, and ``ep_size > 1``. |
| 91 | + """ |
| 92 | + if os.getenv("FASTSAFETENSORS_EP_SLICE", "0") != "1": |
| 93 | + return None |
| 94 | + num_experts = int(os.getenv("FASTSAFETENSORS_EP_NUM_EXPERTS", "0")) |
| 95 | + if num_experts <= 0: |
| 96 | + return None |
| 97 | + ep_size = int(os.getenv("FASTSAFETENSORS_EP_SIZE", "0")) |
| 98 | + ep_rank = int(os.getenv("FASTSAFETENSORS_EP_RANK", "-1")) |
| 99 | + if ep_size <= 0 or ep_rank < 0: |
| 100 | + try: |
| 101 | + import torch.distributed as dist |
| 102 | + |
| 103 | + if dist.is_available() and dist.is_initialized(): |
| 104 | + ep_size = dist.get_world_size() |
| 105 | + ep_rank = dist.get_rank() |
| 106 | + except Exception: |
| 107 | + pass |
| 108 | + if ep_size <= 0: |
| 109 | + ep_size = int(os.getenv("WORLD_SIZE", "1")) |
| 110 | + if ep_rank < 0: |
| 111 | + ep_rank = int(os.getenv("RANK", "0")) |
| 112 | + if ep_size <= 1: |
| 113 | + return None |
| 114 | + return expert_parallel_filter(num_experts, ep_size, ep_rank) |
0 commit comments