| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # SPDX-FileCopyrightText: Songlin Yang, Yu Zhang |
| 4 | # |
| 5 | # This file contains code copied from the flash-linear-attention project. |
| 6 | # The original source code was licensed under the MIT license and included |
| 7 | # the following copyright notice: |
| 8 | # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang |
| 9 | # ruff: noqa: E501 |
| 10 | |
| 11 | |
| 12 | import torch |
| 13 | import torch.nn as nn |
| 14 | |
| 15 | from vllm.model_executor.custom_op import CustomOp |
| 16 | from vllm.triton_utils import tl, triton |
| 17 | from vllm.utils.math_utils import RCP_LN2, cdiv, next_power_of_2 |
| 18 | |
| 19 | from .chunk_delta_h import chunk_gated_delta_rule_fwd_h |
| 20 | from .cumsum import chunk_local_cumsum |
| 21 | from .fused_recurrent import fused_recurrent_gated_delta_rule_fwd_kernel |
| 22 | from .index import prepare_chunk_indices |
| 23 | from .l2norm import l2norm_fwd |
| 24 | from .op import exp2, log |
| 25 | from .solve_tril import solve_tril |
| 26 | from .utils import FLA_CHUNK_SIZE, is_amd |
| 27 | |
| 28 | BT_LIST_AUTOTUNE = [32, 64, 128] |
| 29 | NUM_WARPS_AUTOTUNE = [2, 4, 8, 16] if is_amd else [4, 8, 16, 32] |
| 30 | |
| 31 | |
| 32 | def fused_recurrent_kda_fwd( |
| 33 | q: torch.Tensor, |
| 34 | k: torch.Tensor, |
| 35 | v: torch.Tensor, |
| 36 | g: torch.Tensor, |
| 37 | beta: torch.Tensor, |
| 38 | scale: float, |
| 39 | initial_state: torch.Tensor, |
| 40 | inplace_final_state: bool = True, |
| 41 | cu_seqlens: torch.Tensor | None = None, |
| 42 | ssm_state_indices: torch.Tensor | None = None, |
| 43 | num_accepted_tokens: torch.Tensor | None = None, |
| 44 | use_qk_l2norm_in_kernel: bool = False, |
| 45 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 46 | B, T, H, K, V = *k.shape, v.shape[-1] |
| 47 | HV = v.shape[2] |
| 48 | N = B if cu_seqlens is None else len(cu_seqlens) - 1 |
| 49 | BK, BV = next_power_of_2(K), min(next_power_of_2(V), 8) |
| 50 | NK, NV = cdiv(K, BK), cdiv(V, BV) |
| 51 | assert NK == 1, "NK > 1 is not supported yet" |
| 52 | num_stages = 3 |
| 53 | num_warps = 1 |
| 54 | |
| 55 | o = torch.empty_like(k) |
| 56 | if inplace_final_state: |
| 57 | final_state = initial_state |
| 58 | else: |
| 59 | final_state = q.new_empty(T, HV, V, K, dtype=initial_state.dtype) |
| 60 | |
| 61 | stride_init_state_token = initial_state.stride(0) |
| 62 | stride_final_state_token = final_state.stride(0) |
| 63 | |
| 64 | if ssm_state_indices is None: |
| 65 | stride_indices_seq, stride_indices_tok = 1, 1 |
| 66 | elif ssm_state_indices.ndim == 1: |
| 67 | stride_indices_seq, stride_indices_tok = ssm_state_indices.stride(0), 1 |
| 68 | else: |
| 69 | stride_indices_seq, stride_indices_tok = ssm_state_indices.stride() |
| 70 | |
| 71 | grid = (NK, NV, N * HV) |
| 72 | fused_recurrent_gated_delta_rule_fwd_kernel[grid]( |
| 73 | q=q, |
| 74 | k=k, |
| 75 | v=v, |
| 76 | g=g, |
| 77 | beta=beta, |
| 78 | o=o, |
| 79 | h0=initial_state, |
| 80 | ht=final_state, |
| 81 | cu_seqlens=cu_seqlens, |
| 82 | ssm_state_indices=ssm_state_indices, |
| 83 | num_accepted_tokens=num_accepted_tokens, |
| 84 | scale=scale, |
| 85 | N=N, |
| 86 | T=T, |
| 87 | B=B, |
| 88 | H=H, |
| 89 | HV=HV, |
| 90 | K=K, |
| 91 | V=V, |
| 92 | BK=BK, |
| 93 | BV=BV, |
| 94 | stride_init_state_token=stride_init_state_token, |
| 95 | stride_final_state_token=stride_final_state_token, |
| 96 | stride_indices_seq=stride_indices_seq, |
| 97 | stride_indices_tok=stride_indices_tok, |
| 98 | IS_BETA_HEADWISE=beta.ndim == v.ndim, |
| 99 | USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel, |
| 100 | INPLACE_FINAL_STATE=inplace_final_state, |
| 101 | IS_KDA=True, |
| 102 | num_warps=num_warps, |
| 103 | num_stages=num_stages, |
| 104 | ) |
| 105 | |
| 106 | return o, final_state |
| 107 | |
| 108 | |
| 109 | def fused_recurrent_kda( |
| 110 | q: torch.Tensor, |
| 111 | k: torch.Tensor, |
| 112 | v: torch.Tensor, |
| 113 | g: torch.Tensor, |
| 114 | beta: torch.Tensor = None, |
| 115 | scale: float = None, |
| 116 | initial_state: torch.Tensor = None, |
| 117 | inplace_final_state: bool = True, |
| 118 | use_qk_l2norm_in_kernel: bool = True, |
| 119 | cu_seqlens: torch.Tensor | None = None, |
| 120 | ssm_state_indices: torch.LongTensor | None = None, |
| 121 | **kwargs, |
| 122 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 123 | if cu_seqlens is not None and q.shape[0] != 1: |
| 124 | raise ValueError( |
| 125 | f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`." |
| 126 | f"Please flatten variable-length inputs before processing." |
| 127 | ) |
| 128 | if scale is None: |
| 129 | scale = k.shape[-1] ** -0.5 |
| 130 | |
| 131 | o, final_state = fused_recurrent_kda_fwd( |
| 132 | q=q.contiguous(), |
| 133 | k=k.contiguous(), |
| 134 | v=v.contiguous(), |
| 135 | g=g.contiguous(), |
| 136 | beta=beta.contiguous(), |
| 137 | scale=scale, |
| 138 | initial_state=initial_state, |
| 139 | inplace_final_state=inplace_final_state, |
| 140 | cu_seqlens=cu_seqlens, |
| 141 | ssm_state_indices=ssm_state_indices, |
| 142 | num_accepted_tokens=None, |
| 143 | use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel, |
| 144 | ) |
| 145 | return o, final_state |
| 146 | |
| 147 | |
| 148 | @triton.heuristics( |
| 149 | { |
| 150 | "STORE_RESIDUAL_OUT": lambda args: args["residual_out"] is not None, |
| 151 | "HAS_RESIDUAL": lambda args: args["residual"] is not None, |
| 152 | "HAS_WEIGHT": lambda args: args["w"] is not None, |
| 153 | "HAS_BIAS": lambda args: args["b"] is not None, |
| 154 | } |
| 155 | ) |
| 156 | @triton.jit |
| 157 | def layer_norm_gated_fwd_kernel( |
| 158 | x, # pointer to the input |
| 159 | g, # pointer to the gate |
| 160 | y, # pointer to the output |
| 161 | w, # pointer to the weights |
| 162 | b, # pointer to the biases |
| 163 | residual, # pointer to the residual |
| 164 | residual_out, # pointer to the residual |
| 165 | mean, # pointer to the mean |
| 166 | rstd, # pointer to the 1/std |
| 167 | eps, # epsilon to avoid division by zero |
| 168 | T, # number of rows in x |
| 169 | H: tl.constexpr, # number of heads |
| 170 | g_stride_n: tl.constexpr, |
| 171 | D: tl.constexpr, # number of columns in x |
| 172 | BT: tl.constexpr, |
| 173 | BD: tl.constexpr, |
| 174 | ACTIVATION: tl.constexpr, |
| 175 | IS_RMS_NORM: tl.constexpr, |
| 176 | STORE_RESIDUAL_OUT: tl.constexpr, |
| 177 | HAS_RESIDUAL: tl.constexpr, |
| 178 | HAS_WEIGHT: tl.constexpr, |
| 179 | HAS_BIAS: tl.constexpr, |
| 180 | ): |
| 181 | i_t = tl.program_id(0) |
| 182 | |
| 183 | o_d = tl.arange(0, BD) |
| 184 | m_d = o_d < D |
| 185 | |
| 186 | p_x = tl.make_block_ptr(x, (T, D), (D, 1), (i_t * BT, 0), (BT, BD), (1, 0)) |
| 187 | b_x = tl.load(p_x, boundary_check=(0, 1)).to(tl.float32) |
| 188 | if HAS_RESIDUAL: |
| 189 | p_res = tl.make_block_ptr( |
| 190 | residual, (T, D), (D, 1), (i_t * BT, 0), (BT, BD), (1, 0) |
| 191 | ) |
| 192 | b_x += tl.load(p_res, boundary_check=(0, 1)).to(tl.float32) |
| 193 | if STORE_RESIDUAL_OUT: |
| 194 | p_res_out = tl.make_block_ptr( |
| 195 | residual_out, (T, D), (D, 1), (i_t * BT, 0), (BT, BD), (1, 0) |
| 196 | ) |
| 197 | tl.store(p_res_out, b_x.to(p_res_out.dtype.element_ty), boundary_check=(0, 1)) |
| 198 | if not IS_RMS_NORM: |
| 199 | b_mean = tl.sum(b_x, axis=1) / D |
| 200 | p_mean = tl.make_block_ptr(mean, (T,), (1,), (i_t * BT,), (BT,), (0,)) |
| 201 | tl.store(p_mean, b_mean.to(p_mean.dtype.element_ty), boundary_check=(0,)) |
| 202 | b_xbar = tl.where(m_d[None, :], b_x - b_mean[:, None], 0.0) |
| 203 | b_var = tl.sum(b_xbar * b_xbar, axis=1) / D |
| 204 | else: |
| 205 | b_xbar = tl.where(m_d[None, :], b_x, 0.0) |
| 206 | b_var = tl.sum(b_xbar * b_xbar, axis=1) / D |
| 207 | b_rstd = 1 / tl.sqrt(b_var + eps) |
| 208 | |
| 209 | p_rstd = tl.make_block_ptr(rstd, (T,), (1,), (i_t * BT,), (BT,), (0,)) |
| 210 | tl.store(p_rstd, b_rstd.to(p_rstd.dtype.element_ty), boundary_check=(0,)) |
| 211 | |
| 212 | if HAS_WEIGHT: |
| 213 | b_w = tl.load(w + o_d, mask=m_d).to(tl.float32) |
| 214 | if HAS_BIAS: |
| 215 | b_b = tl.load(b + o_d, mask=m_d).to(tl.float32) |
| 216 | b_x_hat = ( |
| 217 | (b_x - b_mean[:, None]) * b_rstd[:, None] |
| 218 | if not IS_RMS_NORM |
| 219 | else b_x * b_rstd[:, None] |
| 220 | ) |
| 221 | b_y = b_x_hat * b_w[None, :] if HAS_WEIGHT else b_x_hat |
| 222 | if HAS_BIAS: |
| 223 | b_y = b_y + b_b[None, :] |
| 224 | |
| 225 | # swish/sigmoid output gate |
| 226 | o_t = i_t * BT + tl.arange(0, BT) |
| 227 | o_g = (o_t // H) * g_stride_n + (o_t % H) * D |
| 228 | b_g = tl.load( |
| 229 | g + o_g[:, None] + o_d[None, :], |
| 230 | mask=(o_t[:, None] < T) & m_d[None, :], |
| 231 | other=0.0, |
| 232 | ).to(tl.float32) |
| 233 | if ACTIVATION == "swish" or ACTIVATION == "silu": |
| 234 | b_y = b_y * b_g * tl.sigmoid(b_g) |
| 235 | elif ACTIVATION == "sigmoid": |
| 236 | b_y = b_y * tl.sigmoid(b_g) |
| 237 | |
| 238 | # Write output |
| 239 | p_y = tl.make_block_ptr(y, (T, D), (D, 1), (i_t * BT, 0), (BT, BD), (1, 0)) |
| 240 | tl.store(p_y, b_y.to(p_y.dtype.element_ty), boundary_check=(0, 1)) |
| 241 | |
| 242 | |
| 243 | @triton.heuristics( |
| 244 | { |
| 245 | "STORE_RESIDUAL_OUT": lambda args: args["residual_out"] is not None, |
| 246 | "HAS_RESIDUAL": lambda args: args["residual"] is not None, |
| 247 | "HAS_WEIGHT": lambda args: args["w"] is not None, |
| 248 | "HAS_BIAS": lambda args: args["b"] is not None, |
| 249 | } |
| 250 | ) |
| 251 | @triton.jit |
| 252 | def layer_norm_gated_fwd_kernel1( |
| 253 | x, # pointer to the input |
| 254 | g, # pointer to the gate |
| 255 | y, # pointer to the output |
| 256 | w, # pointer to the weights |
| 257 | b, # pointer to the biases |
| 258 | residual, # pointer to the residual |
| 259 | residual_out, # pointer to the residual |
| 260 | mean, # pointer to the mean |
| 261 | rstd, # pointer to the 1/std |
| 262 | eps, # epsilon to avoid division by zero |
| 263 | D: tl.constexpr, # number of columns in x |
| 264 | BD: tl.constexpr, |
| 265 | ACTIVATION: tl.constexpr, |
| 266 | IS_RMS_NORM: tl.constexpr, |
| 267 | STORE_RESIDUAL_OUT: tl.constexpr, |
| 268 | HAS_RESIDUAL: tl.constexpr, |
| 269 | HAS_WEIGHT: tl.constexpr, |
| 270 | HAS_BIAS: tl.constexpr, |
| 271 | ): |
| 272 | i_t = tl.program_id(0) |
| 273 | x += i_t * D |
| 274 | y += i_t * D |
| 275 | g += i_t * D |
| 276 | if HAS_RESIDUAL: |
| 277 | residual += i_t * D |
| 278 | if STORE_RESIDUAL_OUT: |
| 279 | residual_out += i_t * D |
| 280 | |
| 281 | o_d = tl.arange(0, BD) |
| 282 | m_d = o_d < D |
| 283 | b_x = tl.load(x + o_d, mask=m_d, other=0.0).to(tl.float32) |
| 284 | if HAS_RESIDUAL: |
| 285 | b_x += tl.load(residual + o_d, mask=m_d, other=0.0).to(tl.float32) |
| 286 | if STORE_RESIDUAL_OUT: |
| 287 | tl.store(residual_out + o_d, b_x, mask=m_d) |
| 288 | if not IS_RMS_NORM: |
| 289 | b_mean = tl.sum(b_x, axis=0) / D |
| 290 | tl.store(mean + i_t, b_mean) |
| 291 | b_xbar = tl.where(m_d, b_x - b_mean, 0.0) |
| 292 | b_var = tl.sum(b_xbar * b_xbar, axis=0) / D |
| 293 | else: |
| 294 | b_xbar = tl.where(m_d, b_x, 0.0) |
| 295 | b_var = tl.sum(b_xbar * b_xbar, axis=0) / D |
| 296 | b_rstd = 1 / tl.sqrt(b_var + eps) |
| 297 | tl.store(rstd + i_t, b_rstd) |
| 298 | |
| 299 | if HAS_WEIGHT: |
| 300 | b_w = tl.load(w + o_d, mask=m_d).to(tl.float32) |
| 301 | if HAS_BIAS: |
| 302 | b_b = tl.load(b + o_d, mask=m_d).to(tl.float32) |
| 303 | b_x_hat = (b_x - b_mean) * b_rstd if not IS_RMS_NORM else b_x * b_rstd |
| 304 | b_y = b_x_hat * b_w if HAS_WEIGHT else b_x_hat |
| 305 | if HAS_BIAS: |
| 306 | b_y = b_y + b_b |
| 307 | |
| 308 | # swish/sigmoid output gate |
| 309 | b_g = tl.load(g + o_d, mask=m_d, other=0.0).to(tl.float32) |
| 310 | if ACTIVATION == "swish" or ACTIVATION == "silu": |
| 311 | b_y = b_y * b_g * tl.sigmoid(b_g) |
| 312 | elif ACTIVATION == "sigmoid": |
| 313 | b_y = b_y * tl.sigmoid(b_g) |
| 314 | |
| 315 | # Write output |
| 316 | tl.store(y + o_d, b_y, mask=m_d) |
| 317 | |
| 318 | |
| 319 | def layer_norm_gated_fwd( |
| 320 | x: torch.Tensor, |
| 321 | g: torch.Tensor, |
| 322 | weight: torch.Tensor, |
| 323 | bias: torch.Tensor, |
| 324 | activation: str = "swish", |
| 325 | eps: float = 1e-5, |
| 326 | residual: torch.Tensor = None, |
| 327 | out_dtype: torch.dtype = None, |
| 328 | residual_dtype: torch.dtype = None, |
| 329 | is_rms_norm: bool = False, |
| 330 | H: int = 1, |
| 331 | g_stride_n: int | None = None, |
| 332 | ): |
| 333 | if residual is not None: |
| 334 | residual_dtype = residual.dtype |
| 335 | T, D = x.shape |
| 336 | if g_stride_n is None: |
| 337 | g_stride_n = D |
| 338 | assert T % H == 0 |
| 339 | if residual is not None: |
| 340 | assert residual.shape == (T, D) |
| 341 | if weight is not None: |
| 342 | assert weight.shape == (D,) |
| 343 | if bias is not None: |
| 344 | assert bias.shape == (D,) |
| 345 | # allocate output |
| 346 | y = x if out_dtype is None else torch.empty_like(x, dtype=out_dtype) |
| 347 | if residual is not None or ( |
| 348 | residual_dtype is not None and residual_dtype != x.dtype |
| 349 | ): |
| 350 | residual_out = torch.empty(T, D, device=x.device, dtype=residual_dtype) |
| 351 | else: |
| 352 | residual_out = None |
| 353 | mean = ( |
| 354 | torch.empty((T,), dtype=torch.float, device=x.device) |
| 355 | if not is_rms_norm |
| 356 | else None |
| 357 | ) |
| 358 | rstd = torch.empty((T,), dtype=torch.float, device=x.device) |
| 359 | # Less than 64KB per feature: enqueue fused kernel |
| 360 | MAX_FUSED_SIZE = 65536 // x.element_size() |
| 361 | BD = min(MAX_FUSED_SIZE, next_power_of_2(D)) |
| 362 | if D > BD: |
| 363 | raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.") |
| 364 | if D <= 512: |
| 365 | BT = 16 |
| 366 | layer_norm_gated_fwd_kernel[(cdiv(T, BT),)]( |
| 367 | x=x, |
| 368 | g=g, |
| 369 | y=y, |
| 370 | w=weight, |
| 371 | b=bias, |
| 372 | residual=residual, |
| 373 | residual_out=residual_out, |
| 374 | mean=mean, |
| 375 | rstd=rstd, |
| 376 | eps=eps, |
| 377 | T=T, |
| 378 | H=H, |
| 379 | g_stride_n=g_stride_n, |
| 380 | D=D, |
| 381 | BD=BD, |
| 382 | BT=BT, |
| 383 | ACTIVATION=activation, |
| 384 | IS_RMS_NORM=is_rms_norm, |
| 385 | num_warps=8, |
| 386 | ) |
| 387 | else: |
| 388 | layer_norm_gated_fwd_kernel1[(T,)]( |
| 389 | x=x, |
| 390 | g=g, |
| 391 | y=y, |
| 392 | w=weight, |
| 393 | b=bias, |
| 394 | residual=residual, |
| 395 | residual_out=residual_out, |
| 396 | mean=mean, |
| 397 | rstd=rstd, |
| 398 | eps=eps, |
| 399 | D=D, |
| 400 | BD=BD, |
| 401 | ACTIVATION=activation, |
| 402 | IS_RMS_NORM=is_rms_norm, |
| 403 | num_warps=4, |
| 404 | ) |
| 405 | # residual_out is None if residual is None and residual_dtype == input_dtype |
| 406 | return y, mean, rstd, residual_out if residual_out is not None else x |
| 407 | |
| 408 | |
| 409 | def rms_norm_gated( |
| 410 | x: torch.Tensor, |
| 411 | g: torch.Tensor, |
| 412 | weight: torch.Tensor, |
| 413 | bias: torch.Tensor, |
| 414 | activation: str = "swish", |
| 415 | residual: torch.Tensor | None = None, |
| 416 | prenorm: bool = False, |
| 417 | residual_in_fp32: bool = False, |
| 418 | eps: float = 1e-6, |
| 419 | ): |
| 420 | x_shape_og = x.shape |
| 421 | # reshape input data into 2D tensor |
| 422 | x = x.contiguous().reshape(-1, x.shape[-1]) |
| 423 | D = x.shape[-1] |
| 424 | # The tiled kernel supports row-strided gates; kernel1 does not. |
| 425 | if D <= 512: |
| 426 | H = 1 if g.ndim == 2 else g.shape[-2] |
| 427 | g = g.view(-1, H, D) |
| 428 | g_stride_n = g.stride(0) |
| 429 | else: |
| 430 | g = g.contiguous() |
| 431 | H = 1 |
| 432 | g_stride_n = D |
| 433 | if residual is not None: |
| 434 | assert residual.shape == x_shape_og |
| 435 | residual = residual.contiguous().reshape(-1, residual.shape[-1]) |
| 436 | residual_dtype = ( |
| 437 | residual.dtype |
| 438 | if residual is not None |
| 439 | else (torch.float if residual_in_fp32 else None) |
| 440 | ) |
| 441 | y, _, _, residual_out = layer_norm_gated_fwd( |
| 442 | x=x, |
| 443 | g=g, |
| 444 | weight=weight, |
| 445 | bias=bias, |
| 446 | activation=activation, |
| 447 | eps=eps, |
| 448 | residual=residual, |
| 449 | residual_dtype=residual_dtype, |
| 450 | is_rms_norm=True, |
| 451 | H=H, |
| 452 | g_stride_n=g_stride_n, |
| 453 | ) |
| 454 | y = y.reshape(x_shape_og) |
| 455 | return y if not prenorm else (y, residual_out.reshape(x_shape_og)) |
| 456 | |
| 457 | |
| 458 | @CustomOp.register("fused_rms_norm_gated") |
| 459 | class FusedRMSNormGated(CustomOp): |
| 460 | def __init__( |
| 461 | self, |
| 462 | hidden_size: int, |
| 463 | elementwise_affine: bool = True, |
| 464 | eps: float = 1e-5, |
| 465 | activation: str = "swish", |
| 466 | device: torch.device | None = None, |
| 467 | dtype: torch.dtype | None = None, |
| 468 | ) -> None: |
| 469 | factory_kwargs = {"device": device, "dtype": dtype} |
| 470 | super().__init__() |
| 471 | |
| 472 | self.hidden_size = hidden_size |
| 473 | self.elementwise_affine = elementwise_affine |
| 474 | self.eps = eps |
| 475 | self.activation = activation |
| 476 | |
| 477 | if self.activation not in ["swish", "silu", "sigmoid"]: |
| 478 | raise ValueError(f"Unsupported activation: {self.activation}") |
| 479 | |
| 480 | if elementwise_affine: |
| 481 | self.weight = nn.Parameter(torch.empty(hidden_size, **factory_kwargs)) |
| 482 | else: |
| 483 | self.register_parameter("weight", None) |
| 484 | self.register_parameter("bias", None) |
| 485 | |
| 486 | def forward_native( |
| 487 | self, |
| 488 | x: torch.Tensor, |
| 489 | g: torch.Tensor, |
| 490 | residual: torch.Tensor | None = None, |
| 491 | prenorm: bool = False, |
| 492 | residual_in_fp32: bool = False, |
| 493 | ) -> torch.Tensor: |
| 494 | """Decomposed PyTorch ops for torch.compile/inductor fusion.""" |
| 495 | # TODO(https://github.com/vllm-project/vllm/issues/36175): implement |
| 496 | # native residual/prenorm path and unify with RMSNormGated. |
| 497 | # For now, fall back to the triton kernel. |
| 498 | if residual is not None or prenorm: |
| 499 | return self.forward_cuda(x, g, residual, prenorm, residual_in_fp32) |
| 500 | x_float = x.float() |
| 501 | variance = x_float.pow(2).mean(dim=-1, keepdim=True) |
| 502 | x_normed = x_float * torch.rsqrt(variance + self.eps) |
| 503 | if self.weight is not None: |
| 504 | x_normed = x_normed * self.weight.float() |
| 505 | g_float = g.float() |
| 506 | if self.activation in ("swish", "silu"): |
| 507 | out = x_normed * g_float * torch.sigmoid(g_float) |
| 508 | else: # sigmoid |
| 509 | out = x_normed * torch.sigmoid(g_float) |
| 510 | return out.to(x.dtype) |
| 511 | |
| 512 | def forward_cuda( |
| 513 | self, |
| 514 | x: torch.Tensor, |
| 515 | g: torch.Tensor, |
| 516 | residual: torch.Tensor | None = None, |
| 517 | prenorm: bool = False, |
| 518 | residual_in_fp32: bool = False, |
| 519 | ) -> torch.Tensor: |
| 520 | return rms_norm_gated( |
| 521 | x, |
| 522 | g, |
| 523 | self.weight, |
| 524 | self.bias, |
| 525 | self.activation, |
| 526 | residual=residual, |
| 527 | eps=self.eps, |
| 528 | prenorm=prenorm, |
| 529 | residual_in_fp32=residual_in_fp32, |
| 530 | ) |
| 531 | |
| 532 | def forward_xpu( |
| 533 | self, |
| 534 | x: torch.Tensor, |
| 535 | g: torch.Tensor, |
| 536 | residual: torch.Tensor | None = None, |
| 537 | prenorm: bool = False, |
| 538 | residual_in_fp32: bool = False, |
| 539 | ) -> torch.Tensor: |
| 540 | return self.forward_cuda( |
| 541 | x, |
| 542 | g, |
| 543 | residual=residual, |
| 544 | prenorm=prenorm, |
| 545 | residual_in_fp32=residual_in_fp32, |
| 546 | ) |
| 547 | |
| 548 | |
| 549 | @triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None}) |
| 550 | @triton.autotune( |
| 551 | configs=[ |
| 552 | triton.Config({"BK": BK}, num_warps=num_warps, num_stages=num_stages) |
| 553 | for BK in [32, 64] |
| 554 | for num_warps in [1, 2, 4, 8] |
| 555 | for num_stages in [2, 3, 4] |
| 556 | ], |
| 557 | key=["BC"], |
| 558 | ) |
| 559 | @triton.jit(do_not_specialize=["T"]) |
| 560 | def chunk_kda_scaled_dot_kkt_fwd_kernel_intra_sub_inter( |
| 561 | q, |
| 562 | k, |
| 563 | g, |
| 564 | beta, |
| 565 | A, |
| 566 | Aqk, |
| 567 | scale, |
| 568 | cu_seqlens, |
| 569 | chunk_indices, |
| 570 | T, |
| 571 | H: tl.constexpr, |
| 572 | K: tl.constexpr, |
| 573 | BT: tl.constexpr, |
| 574 | BC: tl.constexpr, |
| 575 | BK: tl.constexpr, |
| 576 | NC: tl.constexpr, |
| 577 | IS_VARLEN: tl.constexpr, |
| 578 | ): |
| 579 | i_t, i_c, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2) |
| 580 | i_b, i_h = i_bh // H, i_bh % H |
| 581 | i_i, i_j = i_c // NC, i_c % NC |
| 582 | if IS_VARLEN: |
| 583 | i_n, i_t = ( |
| 584 | tl.load(chunk_indices + i_t * 2).to(tl.int32), |
| 585 | tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32), |
| 586 | ) |
| 587 | bos, eos = ( |
| 588 | tl.load(cu_seqlens + i_n).to(tl.int32), |
| 589 | tl.load(cu_seqlens + i_n + 1).to(tl.int32), |
| 590 | ) |
| 591 | T = eos - bos |
| 592 | else: |
| 593 | bos, eos = i_b * T, i_b * T + T |
| 594 | |
| 595 | if i_t * BT + i_i * BC >= T: |
| 596 | return |
| 597 | if i_i <= i_j: |
| 598 | return |
| 599 | |
| 600 | q += (bos * H + i_h) * K |
| 601 | k += (bos * H + i_h) * K |
| 602 | g += (bos * H + i_h) * K |
| 603 | A += (bos * H + i_h) * BT |
| 604 | Aqk += (bos * H + i_h) * BT |
| 605 | |
| 606 | p_b = tl.make_block_ptr( |
| 607 | beta + bos * H + i_h, (T,), (H,), (i_t * BT + i_i * BC,), (BC,), (0,) |
| 608 | ) |
| 609 | b_b = tl.load(p_b, boundary_check=(0,)) |
| 610 | |
| 611 | b_A = tl.zeros([BC, BC], dtype=tl.float32) |
| 612 | b_Aqk = tl.zeros([BC, BC], dtype=tl.float32) |
| 613 | for i_k in range(tl.cdiv(K, BK)): |
| 614 | p_q = tl.make_block_ptr( |
| 615 | q, (T, K), (H * K, 1), (i_t * BT + i_i * BC, i_k * BK), (BC, BK), (1, 0) |
| 616 | ) |
| 617 | p_k = tl.make_block_ptr( |
| 618 | k, (T, K), (H * K, 1), (i_t * BT + i_i * BC, i_k * BK), (BC, BK), (1, 0) |
| 619 | ) |
| 620 | p_g = tl.make_block_ptr( |
| 621 | g, (T, K), (H * K, 1), (i_t * BT + i_i * BC, i_k * BK), (BC, BK), (1, 0) |
| 622 | ) |
| 623 | b_kt = tl.make_block_ptr( |
| 624 | k, (K, T), (1, H * K), (i_k * BK, i_t * BT + i_j * BC), (BK, BC), (0, 1) |
| 625 | ) |
| 626 | p_gk = tl.make_block_ptr( |
| 627 | g, (K, T), (1, H * K), (i_k * BK, i_t * BT + i_j * BC), (BK, BC), (0, 1) |
| 628 | ) |
| 629 | |
| 630 | o_k = i_k * BK + tl.arange(0, BK) |
| 631 | m_k = o_k < K |
| 632 | # [BK,] |
| 633 | b_gn = tl.load(g + (i_t * BT + i_i * BC) * H * K + o_k, mask=m_k, other=0) |
| 634 | # [BC, BK] |
| 635 | b_g = tl.load(p_g, boundary_check=(0, 1)) |
| 636 | b_k = tl.load(p_k, boundary_check=(0, 1)) * exp2(b_g - b_gn[None, :]) |
| 637 | # [BK, BC] |
| 638 | b_gk = tl.load(p_gk, boundary_check=(0, 1)) |
| 639 | b_kt = tl.load(b_kt, boundary_check=(0, 1)) |
| 640 | # [BC, BC] |
| 641 | b_ktg = b_kt * exp2(b_gn[:, None] - b_gk) |
| 642 | b_A += tl.dot(b_k, b_ktg) |
| 643 | |
| 644 | b_q = tl.load(p_q, boundary_check=(0, 1)) |
| 645 | b_qg = b_q * exp2(b_g - b_gn[None, :]) * scale |
| 646 | b_Aqk += tl.dot(b_qg, b_ktg) |
| 647 | |
| 648 | b_A *= b_b[:, None] |
| 649 | |
| 650 | p_A = tl.make_block_ptr( |
| 651 | A, (T, BT), (H * BT, 1), (i_t * BT + i_i * BC, i_j * BC), (BC, BC), (1, 0) |
| 652 | ) |
| 653 | tl.store(p_A, b_A.to(A.dtype.element_ty), boundary_check=(0, 1)) |
| 654 | p_Aqk = tl.make_block_ptr( |
| 655 | Aqk, (T, BT), (H * BT, 1), (i_t * BT + i_i * BC, i_j * BC), (BC, BC), (1, 0) |
| 656 | ) |
| 657 | tl.store(p_Aqk, b_Aqk.to(Aqk.dtype.element_ty), boundary_check=(0, 1)) |
| 658 | |
| 659 | |
| 660 | @triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None}) |
| 661 | @triton.autotune( |
| 662 | configs=[triton.Config({}, num_warps=num_warps) for num_warps in [1, 2, 4, 8]], |
| 663 | key=["BK", "BT"], |
| 664 | ) |
| 665 | @triton.jit(do_not_specialize=["T"]) |
| 666 | def chunk_kda_scaled_dot_kkt_fwd_kernel_intra_sub_intra( |
| 667 | q, |
| 668 | k, |
| 669 | g, |
| 670 | beta, |
| 671 | A, |
| 672 | Aqk, |
| 673 | scale, |
| 674 | cu_seqlens, |
| 675 | chunk_indices, |
| 676 | T, |
| 677 | H: tl.constexpr, |
| 678 | K: tl.constexpr, |
| 679 | BT: tl.constexpr, |
| 680 | BC: tl.constexpr, |
| 681 | BK: tl.constexpr, |
| 682 | IS_VARLEN: tl.constexpr, |
| 683 | ): |
| 684 | i_t, i_i, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2) |
| 685 | i_b, i_h = i_bh // H, i_bh % H |
| 686 | if IS_VARLEN: |
| 687 | i_n, i_t = ( |
| 688 | tl.load(chunk_indices + i_t * 2).to(tl.int32), |
| 689 | tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32), |
| 690 | ) |
| 691 | bos, eos = ( |
| 692 | tl.load(cu_seqlens + i_n).to(tl.int32), |
| 693 | tl.load(cu_seqlens + i_n + 1).to(tl.int32), |
| 694 | ) |
| 695 | T = eos - bos |
| 696 | else: |
| 697 | bos, eos = i_b * T, i_b * T + T |
| 698 | |
| 699 | if i_t * BT + i_i * BC >= T: |
| 700 | return |
| 701 | |
| 702 | o_i = tl.arange(0, BC) |
| 703 | o_k = tl.arange(0, BK) |
| 704 | m_k = o_k < K |
| 705 | m_A = (i_t * BT + i_i * BC + o_i) < T |
| 706 | o_A = (bos + i_t * BT + i_i * BC + o_i) * H * BT + i_h * BT + i_i * BC |
| 707 | |
| 708 | p_q = tl.make_block_ptr( |
| 709 | q + (bos * H + i_h) * K, |
| 710 | (T, K), |
| 711 | (H * K, 1), |
| 712 | (i_t * BT + i_i * BC, 0), |
| 713 | (BC, BK), |
| 714 | (1, 0), |
| 715 | ) |
| 716 | p_k = tl.make_block_ptr( |
| 717 | k + (bos * H + i_h) * K, |
| 718 | (T, K), |
| 719 | (H * K, 1), |
| 720 | (i_t * BT + i_i * BC, 0), |
| 721 | (BC, BK), |
| 722 | (1, 0), |
| 723 | ) |
| 724 | p_g = tl.make_block_ptr( |
| 725 | g + (bos * H + i_h) * K, |
| 726 | (T, K), |
| 727 | (H * K, 1), |
| 728 | (i_t * BT + i_i * BC, 0), |
| 729 | (BC, BK), |
| 730 | (1, 0), |
| 731 | ) |
| 732 | b_q = tl.load(p_q, boundary_check=(0, 1)) |
| 733 | b_k = tl.load(p_k, boundary_check=(0, 1)) |
| 734 | b_g = tl.load(p_g, boundary_check=(0, 1)) |
| 735 | |
| 736 | p_b = beta + (bos + i_t * BT + i_i * BC + o_i) * H + i_h |
| 737 | b_k = b_k * tl.load(p_b, mask=m_A, other=0)[:, None] |
| 738 | |
| 739 | p_kt = k + (bos + i_t * BT + i_i * BC) * H * K + i_h * K + o_k |
| 740 | p_gk = g + (bos + i_t * BT + i_i * BC) * H * K + i_h * K + o_k |
| 741 | |
| 742 | for j in range(0, min(BC, T - i_t * BT - i_i * BC)): |
| 743 | b_kt = tl.load(p_kt, mask=m_k, other=0).to(tl.float32) |
| 744 | b_gk = tl.load(p_gk, mask=m_k, other=0).to(tl.float32) |
| 745 | b_ktg = b_kt[None, :] * exp2(b_g - b_gk[None, :]) |
| 746 | b_A = tl.sum(b_k * b_ktg, 1) |
| 747 | b_A = tl.where(o_i > j, b_A, 0.0) |
| 748 | b_Aqk = tl.sum(b_q * b_ktg, 1) |
| 749 | b_Aqk = tl.where(o_i >= j, b_Aqk * scale, 0.0) |
| 750 | tl.store(A + o_A + j, b_A, mask=m_A) |
| 751 | tl.store(Aqk + o_A + j, b_Aqk, mask=m_A) |
| 752 | p_kt += H * K |
| 753 | p_gk += H * K |
| 754 | |
| 755 | |
| 756 | def chunk_kda_scaled_dot_kkt_fwd( |
| 757 | q: torch.Tensor, |
| 758 | k: torch.Tensor, |
| 759 | gk: torch.Tensor | None = None, |
| 760 | beta: torch.Tensor | None = None, |
| 761 | scale: float | None = None, |
| 762 | cu_seqlens: torch.Tensor | None = None, |
| 763 | chunk_indices: torch.Tensor | None = None, |
| 764 | chunk_size: int = FLA_CHUNK_SIZE, |
| 765 | output_dtype: torch.dtype = torch.float32, |
| 766 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 767 | r""" |
| 768 | Compute beta * K * K^T. |
| 769 | |
| 770 | Args: |
| 771 | k (torch.Tensor): |
| 772 | The key tensor of shape `[B, T, H, K]`. |
| 773 | beta (torch.Tensor): |
| 774 | The beta tensor of shape `[B, T, H]`. |
| 775 | gk (torch.Tensor): |
| 776 | The cumulative sum of the gate tensor of shape `[B, T, H, K]` applied to the key tensor. Default: `None`. |
| 777 | cu_seqlens (torch.Tensor): |
| 778 | The cumulative sequence lengths of the input tensor. |
| 779 | Default: None |
| 780 | chunk_size (int): |
| 781 | The chunk size. Default: 64. |
| 782 | output_dtype (torch.dtype): |
| 783 | The dtype of the output tensor. Default: `torch.float32` |
| 784 | |
| 785 | Returns: |
| 786 | beta * K * K^T of shape `[B, T, H, BT]` where `BT` is the chunk size. |
| 787 | """ |
| 788 | B, T, H, K = k.shape |
| 789 | assert K <= 256 |
| 790 | BT = chunk_size |
| 791 | if chunk_indices is None and cu_seqlens is not None: |
| 792 | chunk_indices = prepare_chunk_indices(cu_seqlens, BT) |
| 793 | NT = cdiv(T, BT) if cu_seqlens is None else len(chunk_indices) |
| 794 | |
| 795 | BC = min(16, BT) |
| 796 | NC = cdiv(BT, BC) |
| 797 | BK = max(next_power_of_2(K), 16) |
| 798 | A = torch.zeros(B, T, H, BT, device=k.device, dtype=output_dtype) |
| 799 | Aqk = torch.zeros(B, T, H, BT, device=k.device, dtype=output_dtype) |
| 800 | grid = (NT, NC * NC, B * H) |
| 801 | chunk_kda_scaled_dot_kkt_fwd_kernel_intra_sub_inter[grid]( |
| 802 | q=q, |
| 803 | k=k, |
| 804 | g=gk, |
| 805 | beta=beta, |
| 806 | A=A, |
| 807 | Aqk=Aqk, |
| 808 | scale=scale, |
| 809 | cu_seqlens=cu_seqlens, |
| 810 | chunk_indices=chunk_indices, |
| 811 | T=T, |
| 812 | H=H, |
| 813 | K=K, |
| 814 | BT=BT, |
| 815 | BC=BC, |
| 816 | NC=NC, |
| 817 | ) |
| 818 | |
| 819 | grid = (NT, NC, B * H) |
| 820 | chunk_kda_scaled_dot_kkt_fwd_kernel_intra_sub_intra[grid]( |
| 821 | q=q, |
| 822 | k=k, |
| 823 | g=gk, |
| 824 | beta=beta, |
| 825 | A=A, |
| 826 | Aqk=Aqk, |
| 827 | scale=scale, |
| 828 | cu_seqlens=cu_seqlens, |
| 829 | chunk_indices=chunk_indices, |
| 830 | T=T, |
| 831 | H=H, |
| 832 | K=K, |
| 833 | BT=BT, |
| 834 | BC=BC, |
| 835 | BK=BK, |
| 836 | ) |
| 837 | return A, Aqk |
| 838 | |
| 839 | |
| 840 | @triton.heuristics( |
| 841 | { |
| 842 | "STORE_QG": lambda args: args["qg"] is not None, |
| 843 | "STORE_KG": lambda args: args["kg"] is not None, |
| 844 | "IS_VARLEN": lambda args: args["cu_seqlens"] is not None, |
| 845 | } |
| 846 | ) |
| 847 | @triton.autotune( |
| 848 | configs=[ |
| 849 | triton.Config({}, num_warps=num_warps, num_stages=num_stages) |
| 850 | for num_warps in [2, 4, 8] |
| 851 | for num_stages in [2, 3, 4] |
| 852 | ], |
| 853 | key=["H", "K", "V", "BT", "BK", "BV", "IS_VARLEN"], |
| 854 | ) |
| 855 | @triton.jit(do_not_specialize=["T"]) |
| 856 | def recompute_w_u_fwd_kernel( |
| 857 | q, |
| 858 | k, |
| 859 | qg, |
| 860 | kg, |
| 861 | v, |
| 862 | beta, |
| 863 | w, |
| 864 | u, |
| 865 | A, |
| 866 | gk, |
| 867 | cu_seqlens, |
| 868 | chunk_indices, |
| 869 | T, |
| 870 | H: tl.constexpr, |
| 871 | K: tl.constexpr, |
| 872 | V: tl.constexpr, |
| 873 | BT: tl.constexpr, |
| 874 | BK: tl.constexpr, |
| 875 | BV: tl.constexpr, |
| 876 | STORE_QG: tl.constexpr, |
| 877 | STORE_KG: tl.constexpr, |
| 878 | IS_VARLEN: tl.constexpr, |
| 879 | DOT_PRECISION: tl.constexpr, |
| 880 | ): |
| 881 | i_t, i_bh = tl.program_id(0), tl.program_id(1) |
| 882 | i_b, i_h = i_bh // H, i_bh % H |
| 883 | if IS_VARLEN: |
| 884 | i_n, i_t = ( |
| 885 | tl.load(chunk_indices + i_t * 2).to(tl.int32), |
| 886 | tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32), |
| 887 | ) |
| 888 | bos, eos = ( |
| 889 | tl.load(cu_seqlens + i_n).to(tl.int32), |
| 890 | tl.load(cu_seqlens + i_n + 1).to(tl.int32), |
| 891 | ) |
| 892 | T = eos - bos |
| 893 | else: |
| 894 | bos, eos = i_b * T, i_b * T + T |
| 895 | p_b = tl.make_block_ptr(beta + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,)) |
| 896 | b_b = tl.load(p_b, boundary_check=(0,)) |
| 897 | |
| 898 | p_A = tl.make_block_ptr( |
| 899 | A + (bos * H + i_h) * BT, (T, BT), (H * BT, 1), (i_t * BT, 0), (BT, BT), (1, 0) |
| 900 | ) |
| 901 | b_A = tl.load(p_A, boundary_check=(0, 1)) |
| 902 | |
| 903 | for i_v in range(tl.cdiv(V, BV)): |
| 904 | p_v = tl.make_block_ptr( |
| 905 | v + (bos * H + i_h) * V, |
| 906 | (T, V), |
| 907 | (H * V, 1), |
| 908 | (i_t * BT, i_v * BV), |
| 909 | (BT, BV), |
| 910 | (1, 0), |
| 911 | ) |
| 912 | p_u = tl.make_block_ptr( |
| 913 | u + (bos * H + i_h) * V, |
| 914 | (T, V), |
| 915 | (H * V, 1), |
| 916 | (i_t * BT, i_v * BV), |
| 917 | (BT, BV), |
| 918 | (1, 0), |
| 919 | ) |
| 920 | b_v = tl.load(p_v, boundary_check=(0, 1)) |
| 921 | b_vb = (b_v * b_b[:, None]).to(b_v.dtype) |
| 922 | b_u = tl.dot(b_A, b_vb, input_precision=DOT_PRECISION) |
| 923 | tl.store(p_u, b_u.to(p_u.dtype.element_ty), boundary_check=(0, 1)) |
| 924 | |
| 925 | for i_k in range(tl.cdiv(K, BK)): |
| 926 | p_w = tl.make_block_ptr( |
| 927 | w + (bos * H + i_h) * K, |
| 928 | (T, K), |
| 929 | (H * K, 1), |
| 930 | (i_t * BT, i_k * BK), |
| 931 | (BT, BK), |
| 932 | (1, 0), |
| 933 | ) |
| 934 | p_k = tl.make_block_ptr( |
| 935 | k + (bos * H + i_h) * K, |
| 936 | (T, K), |
| 937 | (H * K, 1), |
| 938 | (i_t * BT, i_k * BK), |
| 939 | (BT, BK), |
| 940 | (1, 0), |
| 941 | ) |
| 942 | b_k = tl.load(p_k, boundary_check=(0, 1)) |
| 943 | b_kb = b_k * b_b[:, None] |
| 944 | |
| 945 | p_gk = tl.make_block_ptr( |
| 946 | gk + (bos * H + i_h) * K, |
| 947 | (T, K), |
| 948 | (H * K, 1), |
| 949 | (i_t * BT, i_k * BK), |
| 950 | (BT, BK), |
| 951 | (1, 0), |
| 952 | ) |
| 953 | b_gk = tl.load(p_gk, boundary_check=(0, 1)) |
| 954 | b_kb *= exp2(b_gk) |
| 955 | if STORE_QG: |
| 956 | p_q = tl.make_block_ptr( |
| 957 | q + (bos * H + i_h) * K, |
| 958 | (T, K), |
| 959 | (H * K, 1), |
| 960 | (i_t * BT, i_k * BK), |
| 961 | (BT, BK), |
| 962 | (1, 0), |
| 963 | ) |
| 964 | p_qg = tl.make_block_ptr( |
| 965 | qg + (bos * H + i_h) * K, |
| 966 | (T, K), |
| 967 | (H * K, 1), |
| 968 | (i_t * BT, i_k * BK), |
| 969 | (BT, BK), |
| 970 | (1, 0), |
| 971 | ) |
| 972 | b_q = tl.load(p_q, boundary_check=(0, 1)) |
| 973 | b_qg = b_q * exp2(b_gk) |
| 974 | tl.store(p_qg, b_qg.to(p_qg.dtype.element_ty), boundary_check=(0, 1)) |
| 975 | if STORE_KG: |
| 976 | last_idx = min(i_t * BT + BT, T) - 1 |
| 977 | |
| 978 | o_k = i_k * BK + tl.arange(0, BK) |
| 979 | m_k = o_k < K |
| 980 | b_gn = tl.load( |
| 981 | gk + ((bos + last_idx) * H + i_h) * K + o_k, mask=m_k, other=0.0 |
| 982 | ) |
| 983 | b_kg = b_k * exp2(b_gn - b_gk) |
| 984 | |
| 985 | p_kg = tl.make_block_ptr( |
| 986 | kg + (bos * H + i_h) * K, |
| 987 | (T, K), |
| 988 | (H * K, 1), |
| 989 | (i_t * BT, i_k * BK), |
| 990 | (BT, BK), |
| 991 | (1, 0), |
| 992 | ) |
| 993 | tl.store(p_kg, b_kg.to(p_kg.dtype.element_ty), boundary_check=(0, 1)) |
| 994 | |
| 995 | b_w = tl.dot(b_A, b_kb.to(b_k.dtype)) |
| 996 | tl.store(p_w, b_w.to(p_w.dtype.element_ty), boundary_check=(0, 1)) |
| 997 | |
| 998 | |
| 999 | def recompute_w_u_fwd( |
| 1000 | k: torch.Tensor, |
| 1001 | v: torch.Tensor, |
| 1002 | beta: torch.Tensor, |
| 1003 | A: torch.Tensor, |
| 1004 | q: torch.Tensor | None = None, |
| 1005 | gk: torch.Tensor | None = None, |
| 1006 | cu_seqlens: torch.Tensor | None = None, |
| 1007 | chunk_indices: torch.Tensor | None = None, |
| 1008 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 1009 | B, T, H, K, V = *k.shape, v.shape[-1] |
| 1010 | BT = A.shape[-1] |
| 1011 | BK = 64 |
| 1012 | BV = 64 |
| 1013 | |
| 1014 | if chunk_indices is None and cu_seqlens is not None: |
| 1015 | chunk_indices = prepare_chunk_indices(cu_seqlens, BT) |
| 1016 | NT = cdiv(T, BT) if cu_seqlens is None else len(chunk_indices) |
| 1017 | |
| 1018 | w = torch.empty_like(k) |
| 1019 | u = torch.empty_like(v) |
| 1020 | kg = torch.empty_like(k) if gk is not None else None |
| 1021 | recompute_w_u_fwd_kernel[(NT, B * H)]( |
| 1022 | q=q, |
| 1023 | k=k, |
| 1024 | qg=None, |
| 1025 | kg=kg, |
| 1026 | v=v, |
| 1027 | beta=beta, |
| 1028 | w=w, |
| 1029 | u=u, |
| 1030 | A=A, |
| 1031 | gk=gk, |
| 1032 | cu_seqlens=cu_seqlens, |
| 1033 | chunk_indices=chunk_indices, |
| 1034 | T=T, |
| 1035 | H=H, |
| 1036 | K=K, |
| 1037 | V=V, |
| 1038 | BT=BT, |
| 1039 | BK=BK, |
| 1040 | BV=BV, |
| 1041 | DOT_PRECISION="ieee", |
| 1042 | ) |
| 1043 | return w, u, None, kg |
| 1044 | |
| 1045 | |
| 1046 | @triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None}) |
| 1047 | @triton.autotune( |
| 1048 | configs=[ |
| 1049 | triton.Config({"BK": BK, "BV": BV}, num_warps=num_warps, num_stages=num_stages) |
| 1050 | for BK in [32, 64] |
| 1051 | for BV in [64, 128] |
| 1052 | for num_warps in [2, 4, 8] |
| 1053 | for num_stages in [2, 3, 4] |
| 1054 | ], |
| 1055 | key=["BT"], |
| 1056 | ) |
| 1057 | @triton.jit(do_not_specialize=["T"]) |
| 1058 | def chunk_gla_fwd_kernel_o( |
| 1059 | q, |
| 1060 | v, |
| 1061 | g, |
| 1062 | h, |
| 1063 | o, |
| 1064 | A, |
| 1065 | cu_seqlens, |
| 1066 | chunk_indices, |
| 1067 | scale, |
| 1068 | T, |
| 1069 | H: tl.constexpr, |
| 1070 | K: tl.constexpr, |
| 1071 | V: tl.constexpr, |
| 1072 | BT: tl.constexpr, |
| 1073 | BK: tl.constexpr, |
| 1074 | BV: tl.constexpr, |
| 1075 | IS_VARLEN: tl.constexpr, |
| 1076 | ): |
| 1077 | i_v, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2) |
| 1078 | i_b, i_h = i_bh // H, i_bh % H |
| 1079 | if IS_VARLEN: |
| 1080 | i_tg = i_t |
| 1081 | i_n, i_t = ( |
| 1082 | tl.load(chunk_indices + i_t * 2).to(tl.int32), |
| 1083 | tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32), |
| 1084 | ) |
| 1085 | bos, eos = ( |
| 1086 | tl.load(cu_seqlens + i_n).to(tl.int32), |
| 1087 | tl.load(cu_seqlens + i_n + 1).to(tl.int32), |
| 1088 | ) |
| 1089 | T = eos - bos |
| 1090 | NT = tl.cdiv(T, BT) |
| 1091 | else: |
| 1092 | NT = tl.cdiv(T, BT) |
| 1093 | i_tg = i_b * NT + i_t |
| 1094 | bos, eos = i_b * T, i_b * T + T |
| 1095 | |
| 1096 | m_s = tl.arange(0, BT)[:, None] >= tl.arange(0, BT)[None, :] |
| 1097 | |
| 1098 | b_o = tl.zeros([BT, BV], dtype=tl.float32) |
| 1099 | for i_k in range(tl.cdiv(K, BK)): |
| 1100 | p_q = tl.make_block_ptr( |
| 1101 | q + (bos * H + i_h) * K, |
| 1102 | (T, K), |
| 1103 | (H * K, 1), |
| 1104 | (i_t * BT, i_k * BK), |
| 1105 | (BT, BK), |
| 1106 | (1, 0), |
| 1107 | ) |
| 1108 | p_g = tl.make_block_ptr( |
| 1109 | g + (bos * H + i_h) * K, |
| 1110 | (T, K), |
| 1111 | (H * K, 1), |
| 1112 | (i_t * BT, i_k * BK), |
| 1113 | (BT, BK), |
| 1114 | (1, 0), |
| 1115 | ) |
| 1116 | p_h = tl.make_block_ptr( |
| 1117 | h + (i_tg * H + i_h) * K * V, |
| 1118 | (V, K), |
| 1119 | (K, 1), |
| 1120 | (i_v * BV, i_k * BK), |
| 1121 | (BV, BK), |
| 1122 | (1, 0), |
| 1123 | ) |
| 1124 | |
| 1125 | # [BT, BK] |
| 1126 | b_q = tl.load(p_q, boundary_check=(0, 1)) |
| 1127 | b_q = (b_q * scale).to(b_q.dtype) |
| 1128 | # [BT, BK] |
| 1129 | b_g = tl.load(p_g, boundary_check=(0, 1)) |
| 1130 | # [BT, BK] |
| 1131 | b_qg = (b_q * exp2(b_g)).to(b_q.dtype) |
| 1132 | # [BV, BK] |
| 1133 | b_h = tl.load(p_h, boundary_check=(0, 1)) |
| 1134 | # [BT, BV] |
| 1135 | if i_k >= 0: |
| 1136 | b_o += tl.dot(b_qg, tl.trans(b_h).to(b_qg.dtype)) |
| 1137 | p_v = tl.make_block_ptr( |
| 1138 | v + (bos * H + i_h) * V, |
| 1139 | (T, V), |
| 1140 | (H * V, 1), |
| 1141 | (i_t * BT, i_v * BV), |
| 1142 | (BT, BV), |
| 1143 | (1, 0), |
| 1144 | ) |
| 1145 | p_o = tl.make_block_ptr( |
| 1146 | o + (bos * H + i_h) * V, |
| 1147 | (T, V), |
| 1148 | (H * V, 1), |
| 1149 | (i_t * BT, i_v * BV), |
| 1150 | (BT, BV), |
| 1151 | (1, 0), |
| 1152 | ) |
| 1153 | p_A = tl.make_block_ptr( |
| 1154 | A + (bos * H + i_h) * BT, (T, BT), (H * BT, 1), (i_t * BT, 0), (BT, BT), (1, 0) |
| 1155 | ) |
| 1156 | # [BT, BV] |
| 1157 | b_v = tl.load(p_v, boundary_check=(0, 1)) |
| 1158 | # [BT, BT] |
| 1159 | b_A = tl.load(p_A, boundary_check=(0, 1)) |
| 1160 | b_A = tl.where(m_s, b_A, 0.0).to(b_v.dtype) |
| 1161 | b_o += tl.dot(b_A, b_v, allow_tf32=False) |
| 1162 | tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0, 1)) |
| 1163 | |
| 1164 | |
| 1165 | def chunk_gla_fwd_o_gk( |
| 1166 | q: torch.Tensor, |
| 1167 | v: torch.Tensor, |
| 1168 | g: torch.Tensor, |
| 1169 | A: torch.Tensor, |
| 1170 | h: torch.Tensor, |
| 1171 | o: torch.Tensor, |
| 1172 | scale: float, |
| 1173 | cu_seqlens: torch.Tensor | None = None, |
| 1174 | chunk_indices: torch.Tensor | None = None, |
| 1175 | chunk_size: int = FLA_CHUNK_SIZE, |
| 1176 | ): |
| 1177 | B, T, H, K, V = *q.shape, v.shape[-1] |
| 1178 | BT = chunk_size |
| 1179 | |
| 1180 | if chunk_indices is None and cu_seqlens is not None: |
| 1181 | chunk_indices = prepare_chunk_indices(cu_seqlens, chunk_size) |
| 1182 | NT = cdiv(T, BT) if cu_seqlens is None else len(chunk_indices) |
| 1183 | |
| 1184 | def grid(meta): |
| 1185 | return (cdiv(V, meta["BV"]), NT, B * H) |
| 1186 | |
| 1187 | chunk_gla_fwd_kernel_o[grid]( |
| 1188 | q=q, |
| 1189 | v=v, |
| 1190 | g=g, |
| 1191 | h=h, |
| 1192 | o=o, |
| 1193 | A=A, |
| 1194 | cu_seqlens=cu_seqlens, |
| 1195 | chunk_indices=chunk_indices, |
| 1196 | scale=scale, |
| 1197 | T=T, |
| 1198 | H=H, |
| 1199 | K=K, |
| 1200 | V=V, |
| 1201 | BT=BT, |
| 1202 | ) |
| 1203 | return o |
| 1204 | |
| 1205 | |
| 1206 | @triton.heuristics( |
| 1207 | { |
| 1208 | "HAS_BIAS": lambda args: args["g_bias"] is not None, |
| 1209 | "IS_VARLEN": lambda args: args["cu_seqlens"] is not None, |
| 1210 | } |
| 1211 | ) |
| 1212 | @triton.autotune( |
| 1213 | configs=[ |
| 1214 | triton.Config({"BD": BD}, num_warps=num_warps) |
| 1215 | for BD in [32, 64] |
| 1216 | for num_warps in [2, 4, 8] |
| 1217 | ], |
| 1218 | key=["H", "D", "BT", "IS_VARLEN"], |
| 1219 | ) |
| 1220 | @triton.jit(do_not_specialize=["T"]) |
| 1221 | def kda_gate_cumsum_fwd_kernel( |
| 1222 | g, |
| 1223 | A, |
| 1224 | y, |
| 1225 | g_bias, |
| 1226 | cu_seqlens, |
| 1227 | chunk_indices, |
| 1228 | cumsum_scale, |
| 1229 | lower_bound, |
| 1230 | beta, |
| 1231 | threshold, |
| 1232 | T, |
| 1233 | H: tl.constexpr, |
| 1234 | D: tl.constexpr, |
| 1235 | BT: tl.constexpr, |
| 1236 | BD: tl.constexpr, |
| 1237 | HAS_BIAS: tl.constexpr, |
| 1238 | IS_VARLEN: tl.constexpr, |
| 1239 | USE_LOWER_BOUND: tl.constexpr, |
| 1240 | ): |
| 1241 | i_d, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2) |
| 1242 | i_b, i_h = i_bh // H, i_bh % H |
| 1243 | if IS_VARLEN: |
| 1244 | i_n, i_t = ( |
| 1245 | tl.load(chunk_indices + i_t * 2).to(tl.int32), |
| 1246 | tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32), |
| 1247 | ) |
| 1248 | bos, eos = ( |
| 1249 | tl.load(cu_seqlens + i_n).to(tl.int32), |
| 1250 | tl.load(cu_seqlens + i_n + 1).to(tl.int32), |
| 1251 | ) |
| 1252 | T = eos - bos |
| 1253 | else: |
| 1254 | bos = i_b * T |
| 1255 | |
| 1256 | p_g = tl.make_block_ptr( |
| 1257 | g + (bos * H + i_h) * D, |
| 1258 | (T, D), |
| 1259 | (H * D, 1), |
| 1260 | (i_t * BT, i_d * BD), |
| 1261 | (BT, BD), |
| 1262 | (1, 0), |
| 1263 | ) |
| 1264 | p_y = tl.make_block_ptr( |
| 1265 | y + (bos * H + i_h) * D, |
| 1266 | (T, D), |
| 1267 | (H * D, 1), |
| 1268 | (i_t * BT, i_d * BD), |
| 1269 | (BT, BD), |
| 1270 | (1, 0), |
| 1271 | ) |
| 1272 | |
| 1273 | b_g = tl.load(p_g, boundary_check=(0, 1)).to(tl.float32) |
| 1274 | if HAS_BIAS: |
| 1275 | o_d = i_d * BD + tl.arange(0, BD) |
| 1276 | b_bias = tl.load(g_bias + i_h * D + o_d, mask=o_d < D, other=0.0).to(tl.float32) |
| 1277 | b_g = b_g + b_bias[None, :] |
| 1278 | |
| 1279 | b_a = tl.exp(tl.load(A + i_h).to(tl.float32)) |
| 1280 | if USE_LOWER_BOUND: |
| 1281 | b_gate = lower_bound * tl.sigmoid(b_a * b_g) |
| 1282 | else: |
| 1283 | b_g_scaled = b_g * beta |
| 1284 | b_softplus = tl.where( |
| 1285 | b_g_scaled > threshold, |
| 1286 | b_g, |
| 1287 | (1.0 / beta) * log(1.0 + tl.exp(b_g_scaled)), |
| 1288 | ) |
| 1289 | b_gate = -b_a * b_softplus |
| 1290 | |
| 1291 | # Out-of-bounds rows (load returns 0, but softplus/bias can still make |
| 1292 | # b_gate non-zero) participate in the dot product. They only contribute to |
| 1293 | # out-of-bounds output rows, which are masked away by `boundary_check` on |
| 1294 | # the store, so visible output matches unfused gate + chunk-local cumsum. |
| 1295 | o_t = tl.arange(0, BT) |
| 1296 | m_cumsum = tl.where(o_t[:, None] >= o_t[None, :], 1.0, 0.0) |
| 1297 | b_y = tl.dot(m_cumsum, b_gate, allow_tf32=False) * cumsum_scale |
| 1298 | tl.store(p_y, b_y.to(p_y.dtype.element_ty), boundary_check=(0, 1)) |
| 1299 | |
| 1300 | |
| 1301 | def fused_kda_gate_chunk_cumsum( |
| 1302 | raw_g: torch.Tensor, |
| 1303 | A_log: torch.Tensor, |
| 1304 | g_bias: torch.Tensor | None = None, |
| 1305 | beta: float = 1.0, |
| 1306 | threshold: float = 20.0, |
| 1307 | lower_bound: float | None = None, |
| 1308 | cu_seqlens: torch.Tensor | None = None, |
| 1309 | chunk_indices: torch.Tensor | None = None, |
| 1310 | chunk_size: int = FLA_CHUNK_SIZE, |
| 1311 | output_dtype: torch.dtype | None = torch.float, |
| 1312 | ) -> torch.Tensor: |
| 1313 | if cu_seqlens is not None: |
| 1314 | assert raw_g.shape[0] == 1, ( |
| 1315 | "Only batch size 1 is supported when cu_seqlens are provided" |
| 1316 | ) |
| 1317 | B, T, H, D = raw_g.shape |
| 1318 | if chunk_indices is None and cu_seqlens is not None: |
| 1319 | chunk_indices = prepare_chunk_indices(cu_seqlens, chunk_size) |
| 1320 | NT = cdiv(T, chunk_size) if cu_seqlens is None else len(chunk_indices) |
| 1321 | |
| 1322 | A_log = A_log.reshape(-1) |
| 1323 | if g_bias is not None: |
| 1324 | g_bias = g_bias.reshape(-1) |
| 1325 | y = torch.empty_like(raw_g, dtype=output_dtype or raw_g.dtype) |
| 1326 | |
| 1327 | def grid(meta): |
| 1328 | return (cdiv(meta["D"], meta["BD"]), NT, B * H) |
| 1329 | |
| 1330 | kda_gate_cumsum_fwd_kernel[grid]( |
| 1331 | g=raw_g, |
| 1332 | A=A_log, |
| 1333 | y=y, |
| 1334 | g_bias=g_bias, |
| 1335 | cu_seqlens=cu_seqlens, |
| 1336 | chunk_indices=chunk_indices, |
| 1337 | # RCP_LN2 folds in the natural-log -> log2 conversion so downstream |
| 1338 | # exp2-based kernels reproduce exp(g). Keep this in sync with the |
| 1339 | # `use_exp2=True` path in `_chunk_kda_fwd_with_cumulative_g`. |
| 1340 | cumsum_scale=RCP_LN2, |
| 1341 | lower_bound=lower_bound or 0.0, |
| 1342 | beta=beta, |
| 1343 | threshold=threshold, |
| 1344 | T=T, |
| 1345 | H=H, |
| 1346 | D=D, |
| 1347 | BT=chunk_size, |
| 1348 | USE_LOWER_BOUND=lower_bound is not None, |
| 1349 | ) |
| 1350 | return y |
| 1351 | |
| 1352 | |
| 1353 | def _chunk_kda_fwd_with_cumulative_g( |
| 1354 | q: torch.Tensor, |
| 1355 | k: torch.Tensor, |
| 1356 | v: torch.Tensor, |
| 1357 | g: torch.Tensor, |
| 1358 | beta: torch.Tensor, |
| 1359 | scale: float, |
| 1360 | initial_state: torch.Tensor, |
| 1361 | output_final_state: bool, |
| 1362 | cu_seqlens: torch.Tensor | None = None, |
| 1363 | chunk_indices: torch.Tensor | None = None, |
| 1364 | chunk_size: int = FLA_CHUNK_SIZE, |
| 1365 | ): |
| 1366 | # `g` must already be chunk-local cumulatively-summed AND scaled by |
| 1367 | # RCP_LN2 (so the downstream exp2-based kernels reproduce exp(g)). |
| 1368 | # Use `chunk_kda_fwd` or `chunk_kda_with_fused_gate_fwd` instead of |
| 1369 | # calling this helper directly unless that invariant is upheld. |
| 1370 | # the intra Aqk is kept in fp32 |
| 1371 | # the computation has very marginal effect on the entire throughput |
| 1372 | A, Aqk = chunk_kda_scaled_dot_kkt_fwd( |
| 1373 | q=q, |
| 1374 | k=k, |
| 1375 | gk=g, |
| 1376 | beta=beta, |
| 1377 | scale=scale, |
| 1378 | cu_seqlens=cu_seqlens, |
| 1379 | chunk_indices=chunk_indices, |
| 1380 | chunk_size=chunk_size, |
| 1381 | output_dtype=torch.float32, |
| 1382 | ) |
| 1383 | A = solve_tril(A=A, cu_seqlens=cu_seqlens, output_dtype=k.dtype) |
| 1384 | w, u, _, kg = recompute_w_u_fwd( |
| 1385 | k=k, |
| 1386 | v=v, |
| 1387 | beta=beta, |
| 1388 | A=A, |
| 1389 | gk=g, |
| 1390 | cu_seqlens=cu_seqlens, |
| 1391 | chunk_indices=chunk_indices, |
| 1392 | ) |
| 1393 | del A |
| 1394 | h, v_new, final_state = chunk_gated_delta_rule_fwd_h( |
| 1395 | k=kg, |
| 1396 | w=w, |
| 1397 | u=u, |
| 1398 | gk=g, |
| 1399 | initial_state=initial_state, |
| 1400 | output_final_state=output_final_state, |
| 1401 | cu_seqlens=cu_seqlens, |
| 1402 | chunk_indices=chunk_indices, |
| 1403 | use_exp2=True, |
| 1404 | ) |
| 1405 | del w, u, kg |
| 1406 | o = chunk_gla_fwd_o_gk( |
| 1407 | q=q, |
| 1408 | v=v_new, |
| 1409 | g=g, |
| 1410 | A=Aqk, |
| 1411 | h=h, |
| 1412 | o=v, |
| 1413 | scale=scale, |
| 1414 | cu_seqlens=cu_seqlens, |
| 1415 | chunk_indices=chunk_indices, |
| 1416 | chunk_size=chunk_size, |
| 1417 | ) |
| 1418 | del Aqk, v_new, h |
| 1419 | return o, final_state |
| 1420 | |
| 1421 | |
| 1422 | def chunk_kda_fwd( |
| 1423 | q: torch.Tensor, |
| 1424 | k: torch.Tensor, |
| 1425 | v: torch.Tensor, |
| 1426 | g: torch.Tensor, |
| 1427 | beta: torch.Tensor, |
| 1428 | scale: float, |
| 1429 | initial_state: torch.Tensor, |
| 1430 | output_final_state: bool, |
| 1431 | cu_seqlens: torch.Tensor | None = None, |
| 1432 | ): |
| 1433 | chunk_size = FLA_CHUNK_SIZE |
| 1434 | chunk_indices = ( |
| 1435 | prepare_chunk_indices(cu_seqlens, chunk_size) |
| 1436 | if cu_seqlens is not None |
| 1437 | else None |
| 1438 | ) |
| 1439 | g = chunk_local_cumsum( |
| 1440 | g, |
| 1441 | chunk_size=chunk_size, |
| 1442 | cu_seqlens=cu_seqlens, |
| 1443 | chunk_indices=chunk_indices, |
| 1444 | ) |
| 1445 | # KDA evaluates cumulative gate decays with exp2. Convert from natural-log |
| 1446 | # space so exp(x) is preserved as exp2(x / ln(2)). |
| 1447 | g = g * RCP_LN2 |
| 1448 | return _chunk_kda_fwd_with_cumulative_g( |
| 1449 | q=q, |
| 1450 | k=k, |
| 1451 | v=v, |
| 1452 | g=g, |
| 1453 | beta=beta, |
| 1454 | scale=scale, |
| 1455 | initial_state=initial_state, |
| 1456 | output_final_state=output_final_state, |
| 1457 | cu_seqlens=cu_seqlens, |
| 1458 | chunk_indices=chunk_indices, |
| 1459 | chunk_size=chunk_size, |
| 1460 | ) |
| 1461 | |
| 1462 | |
| 1463 | def chunk_kda_with_fused_gate_fwd( |
| 1464 | q: torch.Tensor, |
| 1465 | k: torch.Tensor, |
| 1466 | v: torch.Tensor, |
| 1467 | raw_g: torch.Tensor, |
| 1468 | beta: torch.Tensor, |
| 1469 | A_log: torch.Tensor, |
| 1470 | g_bias: torch.Tensor | None, |
| 1471 | scale: float, |
| 1472 | initial_state: torch.Tensor, |
| 1473 | output_final_state: bool, |
| 1474 | lower_bound: float | None = None, |
| 1475 | cu_seqlens: torch.Tensor | None = None, |
| 1476 | ): |
| 1477 | chunk_size = FLA_CHUNK_SIZE |
| 1478 | chunk_indices = ( |
| 1479 | prepare_chunk_indices(cu_seqlens, chunk_size) |
| 1480 | if cu_seqlens is not None |
| 1481 | else None |
| 1482 | ) |
| 1483 | g = fused_kda_gate_chunk_cumsum( |
| 1484 | raw_g, |
| 1485 | A_log=A_log, |
| 1486 | g_bias=g_bias, |
| 1487 | cu_seqlens=cu_seqlens, |
| 1488 | chunk_indices=chunk_indices, |
| 1489 | chunk_size=chunk_size, |
| 1490 | lower_bound=lower_bound, |
| 1491 | ) |
| 1492 | return _chunk_kda_fwd_with_cumulative_g( |
| 1493 | q=q, |
| 1494 | k=k, |
| 1495 | v=v, |
| 1496 | g=g, |
| 1497 | beta=beta, |
| 1498 | scale=scale, |
| 1499 | initial_state=initial_state, |
| 1500 | output_final_state=output_final_state, |
| 1501 | cu_seqlens=cu_seqlens, |
| 1502 | chunk_indices=chunk_indices, |
| 1503 | chunk_size=chunk_size, |
| 1504 | ) |
| 1505 | |
| 1506 | |
| 1507 | def chunk_kda( |
| 1508 | q: torch.Tensor, |
| 1509 | k: torch.Tensor, |
| 1510 | v: torch.Tensor, |
| 1511 | g: torch.Tensor, |
| 1512 | beta: torch.Tensor, |
| 1513 | scale: float = None, |
| 1514 | initial_state: torch.Tensor = None, |
| 1515 | output_final_state: bool = False, |
| 1516 | use_qk_l2norm_in_kernel: bool = False, |
| 1517 | cu_seqlens: torch.Tensor | None = None, |
| 1518 | **kwargs, |
| 1519 | ): |
| 1520 | if scale is None: |
| 1521 | scale = k.shape[-1] ** -0.5 |
| 1522 | |
| 1523 | if use_qk_l2norm_in_kernel: |
| 1524 | q = l2norm_fwd(q.contiguous()) |
| 1525 | k = l2norm_fwd(k.contiguous()) |
| 1526 | |
| 1527 | o, final_state = chunk_kda_fwd( |
| 1528 | q=q, |
| 1529 | k=k, |
| 1530 | v=v.contiguous(), |
| 1531 | g=g.contiguous(), |
| 1532 | beta=beta.contiguous(), |
| 1533 | scale=scale, |
| 1534 | initial_state=initial_state.contiguous(), |
| 1535 | output_final_state=output_final_state, |
| 1536 | cu_seqlens=cu_seqlens, |
| 1537 | ) |
| 1538 | return o, final_state |
| 1539 | |
| 1540 | |
| 1541 | def chunk_kda_with_fused_gate( |
| 1542 | q: torch.Tensor, |
| 1543 | k: torch.Tensor, |
| 1544 | v: torch.Tensor, |
| 1545 | raw_g: torch.Tensor, |
| 1546 | beta: torch.Tensor, |
| 1547 | A_log: torch.Tensor, |
| 1548 | g_bias: torch.Tensor | None, |
| 1549 | scale: float | None = None, |
| 1550 | initial_state: torch.Tensor | None = None, |
| 1551 | output_final_state: bool = False, |
| 1552 | lower_bound: float | None = None, |
| 1553 | use_qk_l2norm_in_kernel: bool = False, |
| 1554 | cu_seqlens: torch.Tensor | None = None, |
| 1555 | **kwargs, |
| 1556 | ): |
| 1557 | """Run chunk KDA from raw gate projection using fused gate+cumsum.""" |
| 1558 | if scale is None: |
| 1559 | scale = k.shape[-1] ** -0.5 |
| 1560 | |
| 1561 | if use_qk_l2norm_in_kernel: |
| 1562 | q = l2norm_fwd(q.contiguous()) |
| 1563 | k = l2norm_fwd(k.contiguous()) |
| 1564 | |
| 1565 | o, final_state = chunk_kda_with_fused_gate_fwd( |
| 1566 | q=q, |
| 1567 | k=k, |
| 1568 | v=v.contiguous(), |
| 1569 | raw_g=raw_g.contiguous(), |
| 1570 | beta=beta.contiguous(), |
| 1571 | A_log=A_log, |
| 1572 | g_bias=g_bias, |
| 1573 | scale=scale, |
| 1574 | initial_state=initial_state.contiguous() if initial_state is not None else None, |
| 1575 | output_final_state=output_final_state, |
| 1576 | lower_bound=lower_bound, |
| 1577 | cu_seqlens=cu_seqlens, |
| 1578 | ) |
| 1579 | return o, final_state |
| 1580 | |
| 1581 | |
| 1582 | @triton.autotune( |
| 1583 | configs=[ |
| 1584 | triton.Config({"BT": bt}, num_warps=nw, num_stages=ns) |
| 1585 | for bt in BT_LIST_AUTOTUNE |
| 1586 | for nw in NUM_WARPS_AUTOTUNE |
| 1587 | for ns in [2, 3] |
| 1588 | ], |
| 1589 | key=["H", "D"], |
| 1590 | ) |
| 1591 | @triton.jit |
| 1592 | def kda_gate_fwd_kernel( |
| 1593 | g, |
| 1594 | A, |
| 1595 | y, |
| 1596 | g_bias, |
| 1597 | lower_bound, |
| 1598 | beta: tl.constexpr, |
| 1599 | threshold: tl.constexpr, |
| 1600 | T, |
| 1601 | H, |
| 1602 | D: tl.constexpr, |
| 1603 | BT: tl.constexpr, |
| 1604 | BD: tl.constexpr, |
| 1605 | HAS_BIAS: tl.constexpr, |
| 1606 | USE_LOWER_BOUND: tl.constexpr, |
| 1607 | ): |
| 1608 | i_t, i_h = tl.program_id(0), tl.program_id(1) |
| 1609 | n_t = i_t * BT |
| 1610 | |
| 1611 | b_a = tl.exp(tl.load(A + i_h).to(tl.float32)) |
| 1612 | |
| 1613 | stride_row = H * D |
| 1614 | stride_col = 1 |
| 1615 | |
| 1616 | g_ptr = tl.make_block_ptr( |
| 1617 | base=g + i_h * D, |
| 1618 | shape=(T, D), |
| 1619 | strides=(stride_row, stride_col), |
| 1620 | offsets=(n_t, 0), |
| 1621 | block_shape=(BT, BD), |
| 1622 | order=(1, 0), |
| 1623 | ) |
| 1624 | |
| 1625 | y_ptr = tl.make_block_ptr( |
| 1626 | base=y + i_h * D, |
| 1627 | shape=(T, D), |
| 1628 | strides=(stride_row, stride_col), |
| 1629 | offsets=(n_t, 0), |
| 1630 | block_shape=(BT, BD), |
| 1631 | order=(1, 0), |
| 1632 | ) |
| 1633 | |
| 1634 | b_g = tl.load(g_ptr, boundary_check=(0, 1)).to(tl.float32) |
| 1635 | |
| 1636 | if HAS_BIAS: |
| 1637 | n_d = tl.arange(0, BD) |
| 1638 | bias_mask = n_d < D |
| 1639 | b_bias = tl.load(g_bias + i_h * D + n_d, mask=bias_mask, other=0.0).to( |
| 1640 | tl.float32 |
| 1641 | ) |
| 1642 | b_g = b_g + b_bias[None, :] |
| 1643 | |
| 1644 | if USE_LOWER_BOUND: |
| 1645 | b_y = lower_bound * tl.sigmoid(b_a * b_g) |
| 1646 | else: |
| 1647 | g_scaled = b_g * beta |
| 1648 | use_linear = g_scaled > threshold |
| 1649 | sp = tl.where(use_linear, b_g, (1.0 / beta) * log(1.0 + tl.exp(g_scaled))) |
| 1650 | b_y = -b_a * sp |
| 1651 | |
| 1652 | tl.store(y_ptr, b_y.to(y.dtype.element_ty), boundary_check=(0, 1)) |
| 1653 | |
| 1654 | |
| 1655 | def fused_kda_gate( |
| 1656 | g: torch.Tensor, |
| 1657 | A: torch.Tensor, |
| 1658 | head_k_dim: int, |
| 1659 | g_bias: torch.Tensor | None = None, |
| 1660 | beta: float = 1.0, |
| 1661 | threshold: float = 20.0, |
| 1662 | lower_bound: float | None = None, |
| 1663 | ) -> torch.Tensor: |
| 1664 | """ |
| 1665 | Forward pass for KDA gate: |
| 1666 | input g: [..., H*D] |
| 1667 | param A: [H] or [1, 1, H, 1] |
| 1668 | beta: softplus beta parameter |
| 1669 | threshold: softplus threshold parameter |
| 1670 | return : [..., H, D] |
| 1671 | """ |
| 1672 | orig_shape = g.shape[:-1] |
| 1673 | |
| 1674 | g = g.view(-1, g.shape[-1]) |
| 1675 | T = g.shape[0] |
| 1676 | HD = g.shape[1] |
| 1677 | H = A.numel() |
| 1678 | assert H * head_k_dim == HD |
| 1679 | |
| 1680 | y = torch.empty_like(g, dtype=torch.float32) |
| 1681 | |
| 1682 | def grid(meta): |
| 1683 | return (cdiv(T, meta["BT"]), H) |
| 1684 | |
| 1685 | kda_gate_fwd_kernel[grid]( |
| 1686 | g, |
| 1687 | A, |
| 1688 | y, |
| 1689 | g_bias, |
| 1690 | lower_bound or 0.0, |
| 1691 | beta, |
| 1692 | threshold, |
| 1693 | T, |
| 1694 | H, |
| 1695 | head_k_dim, |
| 1696 | BD=next_power_of_2(head_k_dim), |
| 1697 | HAS_BIAS=g_bias is not None, |
| 1698 | USE_LOWER_BOUND=lower_bound is not None, |
| 1699 | ) |
| 1700 | |
| 1701 | y = y.view(*orig_shape, H, head_k_dim) |
| 1702 | return y |