File size: 12,670 Bytes
c1dc9b2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | """SplitZip support for the ``vmm_fast`` route (native DLL allocator).
Target: ``dynamic_offload.py::_prepare_vmm_fast_runtime`` where
``attach_model_linear_modules(...)`` (ddo-vmm-allocator, ``model_arena_adapter.py``)
returns a ``ModelArenaWeightStore`` whose native shard transport reads EXACTLY
``binding.nbytes`` raw bytes from the safetensors shard straight into the VMM
arena. Python never stages per-weight bytes on that path, so the staging-pool
interception (``splitzip_staging.py``) never sees it.
The seams that make a Python-only integration possible (verified against
ddo-vmm-allocator @ 70802e1, ``model_arena_adapter.py``):
1. All uploads are enqueued on a Python-owned stream:
``self._upload_stream = torch.cuda.Stream(device=device_ordinal)``.
2. ``native.copy_from_host_async(arena_offset, source, stream=...)`` is the
Python-visible DMA used for CPU-source uploads — it copies exactly
``source.nbytes``, so a compressed payload can be DMA'd if it arrives as a
pinned tensor of that size.
3. ``wait_for_compute(ready_event)`` accepts polymorphic tokens, including a
plain ``torch.cuda.Event`` — a decode enqueued after the DMA on the upload
stream is fenced correctly by returning an event recorded after it.
Strategy: replace ``store._native_shard_transport`` with
``SplitZipNativeShardTransport`` AFTER ``attach_model_linear_modules`` returns.
Per compressed binding we read the blob into a persistent pinned buffer, DMA it
into the arena range, decode the full BF16 weight in place on the upload
stream, and return a torch Event. Raw (uncached) bindings are delegated to the
inner transport unchanged. The native DLL is NOT modified.
Because native-shard bindings may carry no retained CPU tensor, cache entries
are additionally aliased to their safetensor slice identity at attach time
(``bind_slice_aliases``) and resolved at fault time via ``binding.file_slice``.
Hardware-verification points are listed in docs/developer/splitzip.md.
"""
from __future__ import annotations
from pathlib import Path
from typing import Any, Callable
import torch
from .splitzip_codec import GpuSplitZipDecoder, PayloadMeta, triton_available
from .splitzip_shard_cache import SplitZipShardCache
# A resolved cache entry: (blob_path, meta, codebook)
_CacheHit = tuple[Any, PayloadMeta, Any]
def _resolve_arena_view(store: Any) -> Callable[[Any, int, int], torch.Tensor]:
"""Find a callable that yields a torch view over arena device memory.
Tries the store, then the native handle. If your build names it
differently, pass ``arena_view_fn`` to :func:`enable_splitzip_fast_transport`.
"""
candidates = []
for owner in (store, getattr(store, "native", None)):
for name in ("arena_view", "arena_view_bytes", "view_arena", "arena_tensor_view"):
fn = getattr(owner, name, None)
if callable(fn):
candidates.append(fn)
if not candidates:
raise RuntimeError(
"SplitZip fast transport could not resolve an arena-view helper on the "
"ModelArenaWeightStore or its native handle. Pass arena_view_fn=... to "
"enable_splitzip_fast_transport (signature: (native, arena_offset, nbytes) "
"-> contiguous torch.uint8 CUDA tensor)."
)
first = candidates[0]
def arena_view_fn(native: Any, arena_offset: int, nbytes: int) -> torch.Tensor:
view = first(native, arena_offset, nbytes)
if not (torch.is_tensor(view) and view.is_cuda and view.is_contiguous()):
raise RuntimeError("arena_view helper returned a non-CUDA or non-contiguous tensor")
return view.reshape(-1) if view.dtype is torch.uint8 else view.reshape(-1).view(torch.uint8)
return arena_view_fn
class SplitZipNativeShardTransport:
"""Wraps the store's ``_NativeShardTransport``; overrides ``upload_batch``.
Contract with ``ModelArenaWeightStore.fault_and_copy_if_needed`` (the only
consumer): called as ``upload_batch([binding], native, upload_stream)`` and
its return value is handed to ``wait_for_compute``. We keep that contract —
including the observed single-binding call pattern. For a mixed multi-
binding batch we raise, because one return token cannot fence both the
inner ring path and our decode; split the batch instead.
"""
def __init__(
self,
inner: Any,
shard_cache: SplitZipShardCache,
store: Any,
*,
arena_view_fn: Callable[[Any, int, int], torch.Tensor] | None = None,
) -> None:
if not triton_available():
raise RuntimeError(
"SplitZip fast transport requires the vendored Triton codec "
"(_splitzip_vendor.codec_gpu); on Windows use the staging-pool route."
)
self._inner = inner
self._shard_cache = shard_cache
self._store = store
self._device_id = getattr(store, "_device_ordinal", 0)
self._decoder = GpuSplitZipDecoder(device_id=self._device_id)
self._arena_view_fn = arena_view_fn or _resolve_arena_view(store)
self._pinned: torch.Tensor | None = None # persistent pinned staging, uint8
self._dec_lut: dict[int, torch.Tensor] = {} # id(codebook) -> GPU dec_lut
self.stats = {"compressed_uploads": 0, "delegated_uploads": 0, "comp_bytes": 0, "raw_bytes": 0}
# Everything except upload_batch delegates to the inner transport (ring
# prefetch, slot management, ...). Prefetched raw reads for compressed
# bindings are wasted work until the adapter learns to skip them; they are
# harmless (the prefetch target is simply not consumed for those bindings).
def __getattr__(self, name: str) -> Any:
return getattr(self._inner, name)
# -- cache lookup -------------------------------------------------------
def _hit_for(self, binding: Any) -> _CacheHit | None:
source = getattr(binding, "source", None)
if source is not None:
hit = self._shard_cache.lookup(source)
if hit is not None:
return hit
# Preferred on the native-shard path: bindings may carry no retained
# CPU tensor, so resolve via the safetensor slice identity aliased at
# attach time (aliasing uses DDO's own offset convention).
file_slice = getattr(binding, "file_slice", None)
if file_slice is not None:
return self._shard_cache.lookup_by_slice(
Path(file_slice.path).name, file_slice.offset_bytes, file_slice.size_bytes
)
return None
def _dec_lut_for(self, codebook: Any, device: torch.device) -> torch.Tensor:
key = id(codebook)
lut = self._dec_lut.get(key)
if lut is None or lut.device != device:
lut = codebook.dec_lut.to(device)
self._dec_lut[key] = lut
return lut
def _pinned_for(self, nbytes: int) -> torch.Tensor:
if self._pinned is None or self._pinned.numel() < nbytes:
self._pinned = torch.empty(nbytes, dtype=torch.uint8, pin_memory=True)
return self._pinned
# -- the seam -----------------------------------------------------------
def upload_batch(self, bindings: list[Any], native: Any, stream: Any) -> Any:
hits = [(b, self._hit_for(b)) for b in bindings]
if not any(hit is not None for _b, hit in hits):
self.stats["delegated_uploads"] += len(bindings)
return self._inner.upload_batch(bindings, native, stream)
if len(bindings) > 1:
raise NotImplementedError(
"SplitZipNativeShardTransport observed a multi-binding upload_batch with "
"compressed content. The adapter's known call pattern is one binding per "
"call; either split the batch or extend this wrapper."
)
binding, hit = hits[0]
return self._upload_compressed(binding, hit, native, stream)
def _upload_compressed(self, binding: Any, hit: _CacheHit, native: Any, stream: Any) -> torch.cuda.Event:
blob_path, meta, codebook = hit
comp = meta.comp_bytes
raw_bytes = meta.n * 2
if comp > raw_bytes:
raise RuntimeError(
f"SplitZip payload exceeds the binding's arena range "
f"({comp} > {raw_bytes}); the coverage gate should have skipped this tensor"
)
pinned = self._pinned_for(comp)
with open(blob_path, "rb", buffering=0) as handle:
read = handle.readinto(memoryview(pinned.numpy()).cast("B")[:comp])
if read != comp:
raise OSError(f"Short SplitZip blob read from {blob_path}: expected {comp}, got {read}")
arena_offset = binding.offset # verified: passed to native.copy_from_host_async
with torch.cuda.stream(stream):
# 1) DMA the compressed payload into the arena range (comp bytes).
native.copy_from_host_async(arena_offset, pinned[:comp], stream=stream)
# 2) Expand bitwise over the same range on the upload stream: the
# decoder first stashes arena[offset:offset+comp] to GPU scratch,
# then writes the full BF16 weight over the arena range.
arena_u8 = self._arena_view_fn(native, arena_offset, raw_bytes)
slot_bf16 = arena_u8.view(torch.bfloat16)
self._decoder.decode_slot_into(
slot_bf16, meta, self._dec_lut_for(codebook, arena_u8.device), stream=stream
)
# 3) Fence: wait_for_compute accepts a plain torch event; it must be
# recorded AFTER the decode so compute never sees compressed bytes.
ready = torch.cuda.Event()
ready.record(stream)
self.stats["compressed_uploads"] += 1
self.stats["comp_bytes"] += comp
self.stats["raw_bytes"] += raw_bytes
return ready
def bind_slice_aliases(root: Any, shard_cache: SplitZipShardCache) -> int:
"""Alias cached content keys to safetensor slice identities, attach time only.
Walks ``root``'s parameters while the loader CPU tensors are alive, reads
each one's ``SafetensorFileSlice`` from the registry DDO bound as
``module._ddo_safetensor_slices``, and records the
(shard name, offset_bytes, size_bytes) triple next to the content key.
After this, ``SplitZipNativeShardTransport`` can resolve compressed
bindings that carry no retained CPU tensor.
Call BEFORE loader tensors are released (i.e. during/just after attach);
returns the number of aliases created.
"""
registry = getattr(root, "_ddo_safetensor_slices", None)
if registry is None:
return 0
lookup = getattr(registry, "lookup", None)
if not callable(lookup):
return 0
aliases = 0
for _name, parameter in root.named_parameters(remove_duplicate=False):
tensor = parameter.data
if tensor is None or tensor.device.type != "cpu":
continue
file_slice = lookup(tensor)
if file_slice is None:
continue
if shard_cache.alias_slice(
tensor, Path(file_slice.path).name, file_slice.offset_bytes, file_slice.size_bytes
):
aliases += 1
return aliases
def enable_splitzip_fast_transport(
store: Any,
shard_cache: SplitZipShardCache,
*,
arena_view_fn: Callable[[Any, int, int], torch.Tensor] | None = None,
) -> SplitZipNativeShardTransport:
"""Swap SplitZip into a ``ModelArenaWeightStore``'s native shard transport.
Call right after ``report, store = attach_model_linear_modules(...)`` in
``_prepare_vmm_fast_runtime`` when ``native_transport_kwargs`` is non-empty
(i.e. the native DLL transport is active), and after
:func:`bind_slice_aliases` while the loader tensors are still alive.
Requires the vendored Triton codec; raises otherwise (the CPU fallback
cannot run on this path because Python never stages per-weight bytes here).
The native DLL is unchanged; this only re-routes file reads for compressed
bindings through the Python-visible DMA seam and adds an in-place GPU
decode on the upload stream.
"""
inner = getattr(store, "_native_shard_transport", None)
if inner is None:
raise RuntimeError(
"store has no _native_shard_transport: native transport is not active. "
"Enable vmm_native_transport, or use the SplitZipStagingPool route instead."
)
wrapper = SplitZipNativeShardTransport(inner, shard_cache, store, arena_view_fn=arena_view_fn)
store._native_shard_transport = wrapper
return wrapper
|