File size: 17,638 Bytes
b025706 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 | # SPDX-FileCopyrightText: © 2023 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import ttnn
from models.common.modules.tt_ccl import get_num_links as get_common_num_links
def get_num_links(mesh_device, cluster_axis=None):
"""
Get the number of available Ethernet links for CCL operations.
This function queries the fabric control plane to determine the maximum number
of usable links for collective communication operations.
Args:
mesh_device: The mesh device to query.
cluster_axis: Optional cluster axis to query links for.
- 0: Query links along the vertical axis (North-South direction).
- 1: Query links along the horizontal axis (East-West direction).
- None: Query links across all axes and return the minimum.
Returns:
int: The number of available links
Example:
>>> num_links = get_num_links(mesh_device)
>>> num_links_axis0 = get_num_links(mesh_device, cluster_axis=0)
"""
return get_common_num_links(mesh_device, cluster_axis)
class TT_CCL:
def __init__(
self,
mesh_device,
):
self.mesh_device = mesh_device
self.sub_device_crs = ttnn.CoreRangeSet(
{
ttnn.CoreRange(
ttnn.CoreCoord(0, 0),
ttnn.CoreCoord(
self.mesh_device.compute_with_storage_grid_size().x - 1,
self.mesh_device.compute_with_storage_grid_size().y - 1,
),
)
}
)
self.barrier_semaphore_idx = [0, 0, 0]
self.barrier_semaphore_handles = [[], [], []]
self.ag_semaphores_idx = [0, 0, 0]
self.ag_semaphore_handles = [[], [], []]
self.rs_semaphores_idx = [0, 0, 0]
self.rs_semaphore_handles = [[], [], []]
# cluster-axis-0, cluster-axis-1, no-cluster-axis
for i in range(3):
# double buffered semaphores
for _ in range(2):
self.barrier_semaphore_handles[i].append(
ttnn.create_global_semaphore(self.mesh_device, self.sub_device_crs, 0)
)
self.ag_semaphore_handles[i].append(
[ttnn.create_global_semaphore(self.mesh_device, self.sub_device_crs, 0) for _ in range(2)]
)
self.rs_semaphore_handles[i].append(
[ttnn.create_global_semaphore(self.mesh_device, self.sub_device_crs, 0) for _ in range(3)]
)
def get_num_links(self, cluster_axis=None):
"""
Get the number of available Ethernet links for CCL operations on this mesh device.
Args:
cluster_axis: Optional cluster axis to query links for.
- 0: Query links along the vertical axis (North-South direction).
- 1: Query links along the horizontal axis (East-West direction).
- None: Query links across all axes and return the minimum.
Returns:
int: The number of available links (minimum 1).
"""
return get_num_links(self.mesh_device, cluster_axis)
# Index 2 stores the no-axis semaphore pool; cluster_axis=0 is a valid axis
# and must not be folded into that bucket.
def get_and_cycle_barrier_semaphore_handle(self, cluster_axis=None):
semaphore_index = 2 if cluster_axis is None else cluster_axis
current_idx = self.barrier_semaphore_idx[semaphore_index]
self.barrier_semaphore_idx[semaphore_index] = (current_idx + 1) % 2
return self.barrier_semaphore_handles[semaphore_index][current_idx]
def get_and_cycle_ag_semaphore_handles(self, cluster_axis=None):
semaphore_index = 2 if cluster_axis is None else cluster_axis
current_idx = self.ag_semaphores_idx[semaphore_index]
self.ag_semaphores_idx[semaphore_index] = (current_idx + 1) % 2
return self.ag_semaphore_handles[semaphore_index][current_idx]
def get_and_cycle_rs_semaphore_handles(self, cluster_axis=None):
semaphore_index = 2 if cluster_axis is None else cluster_axis
current_idx = self.rs_semaphores_idx[semaphore_index]
self.rs_semaphores_idx[semaphore_index] = (current_idx + 1) % 2
return self.rs_semaphore_handles[semaphore_index][current_idx]
def tt_all_reduce(
input_tensor,
mesh_device,
tt_ccl,
cluster_axis=0,
dim=0,
num_reduce_scatter_links=None,
num_all_gather_links=None,
topology=ttnn.Topology.Linear,
memory_config=None,
rs_memory_config=ttnn.DRAM_MEMORY_CONFIG,
sharded=False,
dtype=ttnn.bfloat16,
use_composite=False,
chunks_per_sync=10,
num_workers_per_link=2,
subdevice_id=None,
):
"""
Perform an all-reduce operation across devices in a mesh.
Args:
input_tensor: The input tensor to reduce.
mesh_device: The mesh device to perform the operation on.
tt_ccl: The TT_CCL instance for semaphore management.
cluster_axis: The cluster axis for the reduction (default: 0).
dim: The dimension to reduce along (default: 0).
num_reduce_scatter_links: Number of links for reduce_scatter. If None, uses max available.
num_all_gather_links: Number of links for all_gather. If None, uses max available.
topology: The topology to use (default: ttnn.Topology.Linear).
memory_config: Memory configuration for the output.
sharded: Whether to use sharded memory config.
dtype: Data type for CCL operations.
use_composite: Whether to use composite reduce_scatter + all_gather.
Returns:
The reduced tensor.
"""
# Skip CCL if single device or only 1 device on the target axis
mesh_shape = list(mesh_device.shape)
if mesh_shape == [1, 1] or (cluster_axis == 1 and 1 in list(mesh_device.shape)):
return input_tensor
# Auto-detect num_links if not provided
if num_reduce_scatter_links is None:
num_reduce_scatter_links = tt_ccl.get_num_links(cluster_axis)
if num_all_gather_links is None:
num_all_gather_links = tt_ccl.get_num_links(cluster_axis)
# Ensure dim 0 and 1 are 1
original_shape = input_tensor.shape
if original_shape[0] != 1 or original_shape[1] != 1:
input_tensor = ttnn.reshape(
input_tensor, (1, 1, original_shape[-4] * original_shape[-3] * original_shape[-2], original_shape[-1])
)
# N300 and T3K: reduce_scatter
if 1 in list(mesh_device.shape):
if input_tensor.is_sharded() and not sharded:
input_tensor_sharded = input_tensor
input_tensor = ttnn.sharded_to_interleaved(input_tensor_sharded, ttnn.L1_MEMORY_CONFIG)
input_tensor_sharded.deallocate(True)
reduced = ttnn.experimental.reduce_scatter_minimal_async(
input_tensor,
persistent_output_buffers=None,
dim=dim,
multi_device_global_semaphore=tt_ccl.get_and_cycle_rs_semaphore_handles(),
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(),
num_links=num_reduce_scatter_links,
memory_config=memory_config,
intermediate_memory_config=rs_memory_config,
topology=topology,
chunks_per_sync=chunks_per_sync,
num_workers_per_link=num_workers_per_link,
num_buffers_per_channel=2,
subdevice_id=subdevice_id,
)
input_tensor.deallocate(True)
return reduced
# TG: all_reduce
# Cast to CCL dtype
if input_tensor.dtype != dtype:
input_tensor = ttnn.to_memory_config(input_tensor, ttnn.L1_MEMORY_CONFIG, dtype) # typecast and to interleaved
if sharded and memory_config is not None:
input_tensor = ttnn.to_memory_config(input_tensor, memory_config, dtype) # to sharded
# Ensure the input tensor is in the correct memory configuration
if not sharded: # prefill
input_tensor = ttnn.to_memory_config(input_tensor, ttnn.DRAM_MEMORY_CONFIG)
if not use_composite:
gathered_tensor = ttnn.experimental.all_gather_async(
input_tensor,
persistent_output_buffer=None,
dim=dim,
multi_device_global_semaphore=tt_ccl.get_and_cycle_ag_semaphore_handles(cluster_axis),
num_links=num_all_gather_links,
cluster_axis=cluster_axis,
topology=topology,
memory_config=ttnn.DRAM_MEMORY_CONFIG if not sharded else memory_config,
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(cluster_axis),
chunks_per_sync=10,
num_workers_per_link=2,
num_buffers_per_channel=2,
subdevice_id=subdevice_id,
)
if sharded:
gathered_tensor = ttnn.to_memory_config(gathered_tensor, ttnn.L1_MEMORY_CONFIG)
reduced_tensor = ttnn.experimental.fast_reduce_nc(
gathered_tensor,
dims=[dim],
output=None,
compute_kernel_config=None,
memory_config=ttnn.L1_MEMORY_CONFIG if sharded else ttnn.DRAM_MEMORY_CONFIG,
)
gathered_tensor.deallocate(True)
else:
input_mem_cfg = input_tensor.memory_config()
reduced_tensor = ttnn.experimental.reduce_scatter_minimal_async(
input_tensor,
persistent_output_buffers=None,
dim=dim,
multi_device_global_semaphore=tt_ccl.get_and_cycle_rs_semaphore_handles(cluster_axis),
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(cluster_axis),
num_links=num_reduce_scatter_links,
cluster_axis=cluster_axis,
memory_config=ttnn.DRAM_MEMORY_CONFIG if not sharded else memory_config,
intermediate_memory_config=ttnn.DRAM_MEMORY_CONFIG,
topology=topology,
chunks_per_sync=10,
num_workers_per_link=2,
num_buffers_per_channel=2,
subdevice_id=subdevice_id,
)
reduced_tensor = ttnn.experimental.all_gather_async(
reduced_tensor,
persistent_output_buffer=None,
dim=dim,
multi_device_global_semaphore=tt_ccl.get_and_cycle_ag_semaphore_handles(cluster_axis),
num_links=num_all_gather_links,
cluster_axis=cluster_axis,
topology=topology,
memory_config=input_mem_cfg,
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(cluster_axis),
chunks_per_sync=10,
num_workers_per_link=2,
num_buffers_per_channel=2,
subdevice_id=subdevice_id,
)
# Reshape the reduced tensor to the original shape
reduced_tensor = ttnn.reshape(reduced_tensor, original_shape)
return reduced_tensor
def tt_all_gather(
input_tensor,
mesh_device,
tt_ccl,
cluster_axis,
dim,
num_links=None,
memory_config=None,
sharded=False,
topology=ttnn.Topology.Linear,
dtype=ttnn.bfloat16,
subdevice_id=None,
):
"""
Perform an all-gather operation across devices in a mesh.
Args:
input_tensor: The input tensor to gather.
mesh_device: The mesh device to perform the operation on.
tt_ccl: The TT_CCL instance for semaphore management.
cluster_axis: The cluster axis for the gather operation.
dim: The dimension to gather along.
num_links: Number of links to use. If None, uses max available.
memory_config: Memory configuration for the output.
sharded: Whether to use sharded memory config.
topology: The topology to use (default: ttnn.Topology.Linear).
dtype: Data type for CCL operations.
Returns:
The gathered tensor.
"""
# Skip CCL if single device or only 1 device on the target axis
mesh_shape = list(mesh_device.shape)
if mesh_shape == [1, 1] or (cluster_axis == 1 and 1 in list(mesh_device.shape)):
return input_tensor
# Auto-detect num_links if not provided
if num_links is None:
num_links = tt_ccl.get_num_links(cluster_axis)
# Ensure the input tensor is in the correct memory configuration
if not sharded:
input_tensor = ttnn.to_memory_config(input_tensor, ttnn.DRAM_MEMORY_CONFIG)
# Cast to CCL dtype
if input_tensor.dtype != dtype:
input_tensor = ttnn.to_memory_config(input_tensor, ttnn.L1_MEMORY_CONFIG, dtype) # typecast and to interleaved
if sharded and memory_config is not None:
input_tensor = ttnn.to_memory_config(input_tensor, memory_config, dtype) # to sharded
if cluster_axis is None:
gathered = ttnn.experimental.all_gather_async(
input_tensor,
persistent_output_buffer=None,
dim=dim,
multi_device_global_semaphore=tt_ccl.get_and_cycle_ag_semaphore_handles(),
num_links=num_links,
topology=topology,
memory_config=memory_config,
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(),
chunks_per_sync=10,
num_workers_per_link=2,
num_buffers_per_channel=2,
subdevice_id=subdevice_id,
)
else:
gathered = ttnn.experimental.all_gather_async(
input_tensor,
persistent_output_buffer=None,
dim=dim,
multi_device_global_semaphore=tt_ccl.get_and_cycle_ag_semaphore_handles(cluster_axis),
num_links=num_links,
cluster_axis=cluster_axis,
topology=topology,
memory_config=memory_config,
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(cluster_axis),
chunks_per_sync=10,
num_workers_per_link=2,
num_buffers_per_channel=2,
subdevice_id=subdevice_id,
)
input_tensor.deallocate(True)
return gathered
def tt_distributed_rmsnorm(inp, epsilon, gamma, mesh_device, tt_ccl, compute_kernel_config, num_links=None):
"""
Perform distributed RMS normalization across devices.
Args:
inp: Input tensor.
epsilon: Small value for numerical stability.
gamma: Scale parameter.
mesh_device: The mesh device.
tt_ccl: The TT_CCL instance for semaphore management.
compute_kernel_config: Compute kernel configuration.
num_links: Number of links to use. If None, uses max available for cluster_axis=1.
Returns:
The normalized tensor.
"""
# Auto-detect num_links if not provided
if num_links is None:
num_links = tt_ccl.get_num_links(cluster_axis=1)
# Run distributed rmsnorm part 1
tt_stats = ttnn.rms_norm_pre_all_gather(inp, compute_kernel_config=compute_kernel_config, dtype=ttnn.bfloat16)
padded_shape = (1, 1, inp.shape[-2], 32)
tt_stats = ttnn.reshape(tt_stats, ttnn.Shape(padded_shape)) # TODO: Figure out why we need this
tt_stats_gathered = tt_all_gather(
tt_stats,
mesh_device=mesh_device,
tt_ccl=tt_ccl,
dim=3,
cluster_axis=1,
num_links=num_links,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
)
tt_stats.deallocate(True)
# Run distributed rmsnorm part 2
tt_out = ttnn.rms_norm_post_all_gather(
inp, tt_stats_gathered, epsilon=epsilon, weight=gamma, compute_kernel_config=compute_kernel_config
)
tt_stats_gathered.deallocate(True)
# inp.deallocate(True)
return tt_out
def tt_sharded_distributed_rmsnorm(
inp,
epsilon,
gamma,
mesh_device,
tt_ccl,
ln_sharded_input_memcfg,
ln_sharded_progcfg,
ln_sharded_stats_memcfg,
num_links=None,
):
"""
Perform sharded distributed RMS normalization across devices.
Args:
inp: Input tensor.
epsilon: Small value for numerical stability.
gamma: Scale parameter.
mesh_device: The mesh device.
tt_ccl: The TT_CCL instance for semaphore management.
ln_sharded_input_memcfg: Memory config for sharded input.
ln_sharded_progcfg: Program config for sharded layernorm.
ln_sharded_stats_memcfg: Memory config for sharded stats.
num_links: Number of links to use. If None, uses max available for cluster_axis=1.
Returns:
The normalized tensor.
"""
# Auto-detect num_links if not provided
cluster_axis = 1
if num_links is None:
num_links = tt_ccl.get_num_links(cluster_axis)
inp = ttnn.to_memory_config(inp, memory_config=ln_sharded_input_memcfg)
# Run distributed rmsnorm part 1
tt_stats = ttnn.rms_norm_pre_all_gather(inp, program_config=ln_sharded_progcfg)
# All gather stats
tt_stats = ttnn.experimental.all_gather_async(
tt_stats,
persistent_output_buffer=None,
dim=3,
multi_device_global_semaphore=tt_ccl.get_and_cycle_ag_semaphore_handles(cluster_axis),
num_links=num_links,
cluster_axis=cluster_axis,
topology=ttnn.Topology.Linear,
memory_config=ln_sharded_stats_memcfg,
barrier_semaphore=tt_ccl.get_and_cycle_barrier_semaphore_handle(cluster_axis),
chunks_per_sync=10,
num_workers_per_link=2,
num_buffers_per_channel=2,
)
# Run distributed rmsnorm part 2
tt_out = ttnn.rms_norm_post_all_gather(
inp,
epsilon=epsilon,
weight=gamma,
program_config=ln_sharded_progcfg,
stats=tt_stats,
)
tt_stats.deallocate(True)
return tt_out
|