File size: 14,030 Bytes
e649a04 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 | # SPDX-FileCopyrightText: © 2026 Tenstorrent AI ULC
# SPDX-License-Identifier: Apache-2.0
"""Prefill host/device input staging and replay refresh."""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Callable, Sequence
import torch
import ttnn
from models.common.llm_runtime.prefill.plan import PrefillChunk, PrefillRequest
from models.common.llm_runtime.prefill.sampling_helpers import _TILE_SIZE
from models.common.llm_runtime.tensor_resources import (
attach_cleanup_failures,
best_effort_deallocate_owned_tensors,
raise_cleanup_failures,
)
@dataclass(frozen=True)
class PrefillHostInputs:
tokens: Any
position_indices: Any
page_table: Any
chunk_page_table: Any | None
chunk_start_idx: Any | None
def values(self) -> tuple[Any, ...]:
return (
self.tokens,
self.position_indices,
self.page_table,
self.chunk_page_table,
self.chunk_start_idx,
)
@dataclass(frozen=True)
class PrefillDeviceInputs:
tokens: Any
rotary_cos: Any
rotary_sin: Any
page_table: Any
chunk_page_table: Any | None
position_indices: Any
chunk_start_idx: Any | None
def model_values(self) -> tuple[Any, ...]:
return (
self.tokens,
self.rotary_cos,
self.rotary_sin,
self.page_table,
self.chunk_page_table,
self.position_indices,
self.chunk_start_idx,
)
def owned_tensor_values(self) -> tuple[Any, ...]:
return self.model_values()
@dataclass(frozen=True)
class PrefillPositionInputs:
slice_start: Any
slice_end: Any
row_index: Any
def values(self) -> tuple[Any, ...]:
return self.slice_start, self.slice_end, self.row_index
def owned_tensor_values(self) -> tuple[Any, ...]:
return self.values()
@dataclass(frozen=True)
class PrefillTraceInputs:
"""Named geometry for allocating one persistent trace input set."""
chunk: PrefillChunk
relative_last: int
tokens: torch.Tensor
start_pos: int
chunk_page_table: torch.Tensor | None
chunk_start_idx: int | None
sequence_length: int
class PrefillInputStager:
"""Own TT prefill tensor geometry, allocation, and in-place refresh."""
def __init__(
self,
*,
model: Any,
mesh_device: Any,
release_transient: Callable[[Any], list[BaseException]],
) -> None:
self.model = model
self.mesh_device = mesh_device
self._release_transient = release_transient
def trace_inputs(self, request: PrefillRequest) -> PrefillTraceInputs:
chunk = request.chunks[0]
if request.uses_chunked_prefill:
final_chunk = request.chunks[-1]
relative_last = (request.last_token_indices[0] - final_chunk.chunk_start_idx) % final_chunk.chunk_size
tokens = request.tokens[:, chunk.token_slice]
start_pos = chunk.chunk_start_idx
chunk_page_table = chunk.chunk_page_table
chunk_start_idx = chunk.chunk_start_idx
sequence_length = chunk.chunk_size
else:
relative_last = max(
last - cached for last, cached in zip(request.last_token_indices, request.cached_tokens)
)
tokens = request.tokens
start_pos = 0
chunk_page_table = None
chunk_start_idx = None
sequence_length = request.padded_sequence_length
return PrefillTraceInputs(
chunk=chunk,
relative_last=relative_last,
tokens=tokens,
start_pos=start_pos,
chunk_page_table=chunk_page_table,
chunk_start_idx=chunk_start_idx,
sequence_length=sequence_length,
)
def stage_step(
self,
request: PrefillRequest,
chunk: PrefillChunk,
final_relative_last: int,
) -> tuple[PrefillDeviceInputs, PrefillPositionInputs]:
chunked = request.uses_chunked_prefill
host_inputs = self.prepare_host_inputs(
request.tokens[:, chunk.token_slice],
request.page_table,
start_pos=chunk.chunk_start_idx if chunked else 0,
chunk_page_table=chunk.chunk_page_table if chunked else None,
chunk_start_idx=chunk.chunk_start_idx if chunked else None,
last_token_idx=max(request.last_token_indices),
)
device_inputs = None
position_inputs = None
try:
device_inputs = self.stage_device_inputs(host_inputs)
position_values = allocate_device_tensors(
self.prepare_position_inputs_host(final_relative_last, chunk.chunk_size).values(),
mesh_device=self.mesh_device,
)
position_inputs = PrefillPositionInputs(*position_values)
except BaseException as primary:
failures = self._release_transient((device_inputs, position_inputs))
attach_cleanup_failures(primary, failures)
raise
return device_inputs, position_inputs
def prepare_host_inputs(
self,
tokens: torch.Tensor,
page_table: torch.Tensor,
*,
start_pos: int = 0,
chunk_page_table: torch.Tensor | None = None,
chunk_start_idx: int | None = None,
last_token_idx: int | None = None,
) -> PrefillHostInputs:
if tokens.ndim != 2:
raise ValueError("prefill tokens must be rank 2")
mapper = ttnn.ReplicateTensorToMesh(self.mesh_device)
tokens_tt = ttnn.from_torch(
tokens.reshape(1, 1, 1, -1),
device=None,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
rope = self.model.rope_setup
rope.load_device_weights()
matrix_length = int(rope.cos_matrix.shape[2])
if matrix_length <= 0:
raise ValueError("rotary position table must not be empty")
start_pos = int(start_pos)
sequence_length = int(tokens.shape[-1])
if start_pos < 0:
raise ValueError("prefill start position must be nonnegative")
if last_token_idx is not None and int(last_token_idx) + 1 > matrix_length:
raise ValueError(f"Sequence length {int(last_token_idx) + 1} exceeds rotary capacity {matrix_length}")
position_indices = torch.arange(start_pos, start_pos + sequence_length, dtype=torch.long).clamp(
max=matrix_length - 1
)
position_indices_tt = ttnn.from_torch(
position_indices.reshape(1, -1),
device=None,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
page_table_tt = ttnn.from_torch(
page_table,
device=None,
dtype=ttnn.int32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
chunk_tt = (
ttnn.from_torch(
chunk_page_table,
device=None,
dtype=ttnn.int32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
if chunk_page_table is not None
else None
)
chunk_start_tt = (
ttnn.from_torch(
torch.tensor([int(chunk_start_idx)], dtype=torch.int32),
device=None,
dtype=ttnn.int32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
if chunk_start_idx is not None
else None
)
return PrefillHostInputs(tokens_tt, position_indices_tt, page_table_tt, chunk_tt, chunk_start_tt)
def prepare_position_inputs_host(self, relative_last: int, sequence_length: int) -> PrefillPositionInputs:
relative_last = int(relative_last)
sequence_length = int(sequence_length)
if relative_last < 0 or relative_last >= sequence_length:
raise ValueError("prefill last-token position must fall within the padded sequence")
block_start = (relative_last // _TILE_SIZE) * _TILE_SIZE
hidden_width = int(self.model.config.dim)
bounds = ((0, 0, block_start, 0), (1, 1, block_start + _TILE_SIZE, hidden_width))
mapper = ttnn.ReplicateTensorToMesh(self.mesh_device)
slice_bounds = tuple(
ttnn.from_torch(
torch.tensor(bound, dtype=torch.int32),
device=None,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
for bound in bounds
)
row_index = ttnn.from_torch(
torch.tensor([[relative_last % _TILE_SIZE]], dtype=torch.int32),
device=None,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
return PrefillPositionInputs(slice_bounds[0], slice_bounds[1], row_index)
def stage_device_inputs(self, host_inputs: PrefillHostInputs) -> PrefillDeviceInputs:
raw_inputs = None
rot_mats = None
try:
raw_inputs = allocate_device_tensors(host_inputs.values(), mesh_device=self.mesh_device)
prepare_rot_mats = getattr(self.model, "prepare_prefill_rot_mats", None)
if not callable(prepare_rot_mats):
raise TypeError("model must provide prepare_prefill_rot_mats()")
rot_mats = tuple(prepare_rot_mats(raw_inputs[1]))
if len(rot_mats) != 2:
raise ValueError("prepare_prefill_rot_mats() must return cosine and sine tensors")
except BaseException as primary:
failures = self._release_transient((rot_mats, raw_inputs))
attach_cleanup_failures(primary, failures)
raise
return PrefillDeviceInputs(
tokens=raw_inputs[0],
rotary_cos=rot_mats[0],
rotary_sin=rot_mats[1],
page_table=raw_inputs[2],
chunk_page_table=raw_inputs[3],
position_indices=raw_inputs[1],
chunk_start_idx=raw_inputs[4],
)
def refresh_regular_device_inputs(self, request: PrefillRequest, device_inputs: PrefillDeviceInputs) -> None:
mapper = ttnn.ReplicateTensorToMesh(self.mesh_device)
host_tokens = ttnn.from_torch(
request.tokens.reshape(1, 1, 1, -1),
device=None,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
host_page_table = ttnn.from_torch(
request.page_table,
device=None,
dtype=ttnn.int32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
copy_into_device_tensors((host_tokens, host_page_table), (device_inputs.tokens, device_inputs.page_table))
def refresh_chunk_device_inputs(
self,
request: PrefillRequest,
chunk: PrefillChunk,
device_inputs: PrefillDeviceInputs,
) -> None:
host_inputs = self.prepare_host_inputs(
request.tokens[:, chunk.token_slice],
request.page_table,
start_pos=chunk.chunk_start_idx,
chunk_page_table=chunk.chunk_page_table,
chunk_start_idx=chunk.chunk_start_idx,
last_token_idx=max(request.last_token_indices),
)
copy_into_device_tensors(
host_inputs.values(),
(
device_inputs.tokens,
device_inputs.position_indices,
device_inputs.page_table,
device_inputs.chunk_page_table,
device_inputs.chunk_start_idx,
),
)
self.copy_rotary_inputs(device_inputs)
def copy_rotary_inputs(self, device_inputs: PrefillDeviceInputs) -> None:
rot_mats = None
try:
rot_mats = tuple(self.model.prepare_prefill_rot_mats(device_inputs.position_indices))
if len(rot_mats) != 2:
raise ValueError("prepare_prefill_rot_mats() must return cosine and sine tensors")
ttnn.copy(input_a=rot_mats[0], input_b=device_inputs.rotary_cos)
ttnn.copy(input_a=rot_mats[1], input_b=device_inputs.rotary_sin)
except BaseException as primary:
failures = self._release_transient(rot_mats)
attach_cleanup_failures(primary, failures)
raise
failures = self._release_transient(rot_mats)
if failures:
raise_cleanup_failures(failures)
def allocate_device_tensors(host_tensors: Sequence[Any], *, mesh_device: Any) -> list[Any]:
"""Allocate device tensors corresponding to one host tensor structure."""
allocated = []
try:
for host_tensor in host_tensors:
allocated.append(ttnn.to_device(host_tensor, device=mesh_device) if host_tensor is not None else None)
except BaseException as primary:
failures = best_effort_deallocate_owned_tensors(allocated)
attach_cleanup_failures(primary, failures)
raise
return allocated
def copy_into_device_tensors(host_tensors: Sequence[Any], device_tensors: Sequence[Any]) -> Sequence[Any]:
"""Refresh an existing device tensor structure without allocating it."""
if len(host_tensors) != len(device_tensors):
raise ValueError("host/device tensor structures must have equal length")
if any(
(host_tensor is None) != (device_tensor is None)
for host_tensor, device_tensor in zip(host_tensors, device_tensors)
):
raise ValueError("host/device optional tensor structure changed")
for host_tensor, device_tensor in zip(host_tensors, device_tensors):
if host_tensor is None:
continue
ttnn.copy_host_to_device_tensor(host_tensor, device_tensor)
return device_tensors
|