File size: 9,670 Bytes
2415c4c | 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 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
"""
TTTv2-style Embedding module for 1D-topology devices: N150 (1x1), N300 (1x2), T3K (1x8).
Single unified Embedding1D class:
- forward(x): Token embedding lookup, optionally scaled by embed_scale.
Execution path:
embedding(x, weights) → [multiply(embed_scale) if embed_scale != 1.0]
This module replaces both TTTv1 Embedding and ScaledEmbedding classes.
"""
from dataclasses import dataclass, replace
from pathlib import Path
import ttnn
from models.common.lightweightmodule import LightweightModule
from models.common.modules.lazy_weight import LazyWeight, resolve_lazy_weight
# =============================================================================
# Top-level config dataclass
# =============================================================================
@dataclass
class Embedding1DConfig:
"""
Central configuration for Embedding1D - the single source of truth for all settings.
Simple usage (all defaults):
config = Embedding1DConfig(weights=lazy_weights)
With scale:
config = Embedding1DConfig(weights=lazy_weights, embed_scale=3072**0.5)
Full customization:
config = Embedding1DConfig(
weights=lazy_weights,
mesh_device=custom_device,
weights_memcfg=custom_memcfg,
)
"""
# Required: embedding weight table (vocab_size, dim), sharded on dim=-1
weights: LazyWeight
# Optional: device (derived from weights if None)
mesh_device: ttnn.MeshDevice | None = None
# Optional: scaling factor applied after lookup (1.0 = no scaling)
embed_scale: float = 1.0
# Optional: power-user overrides (None = compute defaults)
weights_dtype: ttnn.DataType | None = None
weights_memcfg: ttnn.MemoryConfig | None = None
output_memcfg: ttnn.MemoryConfig | None = None
def is_resolved(self) -> bool:
"""Check if all fields are resolved."""
return all(getattr(self, f) is not None for f in self.__dataclass_fields__)
# =============================================================================
# Embedding1D - Unified Embedding for 1D-topology devices
# =============================================================================
class Embedding1D(LightweightModule):
"""
Embedding for non-TG devices supporting token lookup with optional scaling.
Replaces both TTTv1 Embedding and ScaledEmbedding.
Simple API (90% of users):
emb = Embedding1D(weights)
With scaling:
emb = Embedding1D(weights, embed_scale=3072**0.5)
Power API (10% of users):
config = Embedding1DConfig(weights=lazy_w, weights_memcfg=custom_cfg)
emb = Embedding1D.from_config(config)
Execution path:
embedding(x, weights) → [multiply(embed_scale) if embed_scale != 1.0]
"""
def __init__(self, weights: LazyWeight, embed_scale: float = 1.0):
"""
Simple API - derives all config from weights.
Args:
weights: Embedding weight table (vocab_size, dim), sharded on last dim.
embed_scale: Scale factor applied after lookup (default 1.0 = no scaling).
"""
super().__init__()
self.config = _resolve_embedding1d_config(Embedding1DConfig(weights=weights, embed_scale=embed_scale))
self._device_weights_loaded = False
@classmethod
def from_config(cls, config: Embedding1DConfig):
"""
Power API - any level of customization via config.
Override any subset of fields in Embedding1DConfig:
config = Embedding1DConfig(weights=w, embed_scale=2.0)
emb = Embedding1D.from_config(config)
"""
instance = object.__new__(cls)
super(Embedding1D, instance).__init__()
instance.config = _resolve_embedding1d_config(config)
instance._device_weights_loaded = False
return instance
def load_device_weights(self):
if self._device_weights_loaded:
return
assert self.config.is_resolved(), "config must be resolved before loading device weights!"
self.weights = self.config.weights.get_device_weight()
self._device_weights_loaded = True
def forward(self, x: ttnn.Tensor | LazyWeight) -> ttnn.Tensor:
"""
Token embedding lookup.
Args:
x: Token IDs tensor (uint32), shape [1, 1, 1, seq_len].
Returns:
Embedded tokens, shape [1, 1, seq_len, dim].
"""
self.load_device_weights()
x = _load_input_device_tensor(x, self.config)
out = ttnn.embedding(
x,
self.weights,
layout=ttnn.TILE_LAYOUT,
memory_config=self.config.output_memcfg,
)
if self.config.embed_scale != 1.0:
out = ttnn.multiply(out, self.config.embed_scale, memory_config=self.config.output_memcfg)
return out
# [INFO] this is the entry point for TTTv1 model_config.py and will retire with TTTv1
@classmethod
def from_model_args(
cls,
mesh_device,
args,
weight_cache_path,
state_dict,
dtype,
embed_scale: float = 1.0,
):
"""Factory method for backward compatibility with ModelArgs.
Args:
mesh_device: The mesh device to use.
args: Model arguments (ModelArgs instance).
weight_cache_path: Path for weight caching.
state_dict: The state dictionary containing weights.
dtype: Data type for weights.
embed_scale: Scale factor applied after lookup (default 1.0).
"""
if args.is_galaxy:
raise ValueError("Embedding1D cannot be used for Galaxy devices.")
base_name = args.get_state_dict_prefix("", None) + "tok_embeddings.weight"
torch_weight = state_dict[base_name].unsqueeze(0).unsqueeze(0)
cache_dir = None if args.dummy_weights else Path(weight_cache_path) / "embedding"
num_devices = mesh_device.get_num_devices()
weights = LazyWeight(
source=torch_weight,
dtype=dtype,
device=mesh_device,
mesh_mapper_config=ttnn.MeshMapperConfig(
placements=[ttnn.PlacementShard(-1)],
mesh_shape_override=ttnn.MeshShape([num_devices]),
),
layout=ttnn.ROW_MAJOR_LAYOUT,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
cache_dir_weight_name=(cache_dir, "tok_embeddings") if cache_dir else None,
)
config = Embedding1DConfig(
weights=weights,
mesh_device=mesh_device,
embed_scale=embed_scale,
weights_dtype=dtype,
weights_memcfg=ttnn.DRAM_MEMORY_CONFIG,
output_memcfg=ttnn.DRAM_MEMORY_CONFIG,
)
return cls.from_config(config)
# =============================================================================
# Config resolution
# =============================================================================
def _resolve_embedding1d_config(config: Embedding1DConfig) -> Embedding1DConfig:
"""Materialize the config to known good defaults using replace pattern."""
to_set = {}
# --- Phase 1: Foundational fields ---
# Derive mesh_device from weights
mesh_device = config.mesh_device
if mesh_device is None:
mesh_device = config.weights.device
if mesh_device is None:
mesh_device = ttnn.GetDefaultDevice()
if config.mesh_device is None:
to_set["mesh_device"] = mesh_device
assert mesh_device is not None, "mesh_device must be available at this point!"
# --- Phase 2: Default configs ---
if config.weights_dtype is None:
to_set["weights_dtype"] = ttnn.bfloat16
if config.weights_memcfg is None:
to_set["weights_memcfg"] = ttnn.DRAM_MEMORY_CONFIG
if config.output_memcfg is None:
to_set["output_memcfg"] = ttnn.DRAM_MEMORY_CONFIG
# --- Phase 3: Resolve LazyWeight ---
num_devices = mesh_device.get_num_devices()
weights_memcfg = (
config.weights_memcfg
if config.weights_memcfg is not None
else to_set.get("weights_memcfg", ttnn.DRAM_MEMORY_CONFIG)
)
weights_dtype = (
config.weights_dtype if config.weights_dtype is not None else to_set.get("weights_dtype", ttnn.bfloat16)
)
to_set["weights"] = resolve_lazy_weight(
config.weights,
device=mesh_device,
memory_config=weights_memcfg,
mesh_mapper_config=ttnn.MeshMapperConfig(
placements=[ttnn.PlacementShard(-1)],
mesh_shape_override=ttnn.MeshShape([num_devices]),
),
layout=ttnn.ROW_MAJOR_LAYOUT,
dtype=weights_dtype,
)
resolved_config = replace(config, **to_set)
assert resolved_config.is_resolved(), "Config must be resolved!"
assert resolved_config.weights.is_resolved(), "Weights must be resolved!"
return resolved_config
def _load_input_device_tensor(x: ttnn.Tensor | LazyWeight, config: Embedding1DConfig) -> ttnn.Tensor:
"""Resolve the input tensor to ttnn tensor if x is a LazyWeight, otherwise return as-is."""
if isinstance(x, LazyWeight):
resolved_x = resolve_lazy_weight(
x,
device=config.mesh_device,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper_config=None, # replicated
layout=ttnn.ROW_MAJOR_LAYOUT,
)
return resolved_x.get_device_weight()
assert isinstance(x, ttnn.Tensor), "x must be a ttnn tensor at this point!"
return x
|