| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Fused Qwen4Exp PLE kernels. |
| 4 | |
| 5 | N-gram IDs hash each suffix with |
| 6 | ``offset[h] + (xor_i(token[t-i] * multiplier[i]) % size[h])``. The gate computes |
| 7 | ``d = dot(RMSNorm(key), RMSNorm(query)) / sqrt(H)`` and |
| 8 | ``g = sigmoid(sign(d) * sqrt(max(abs(d), 1e-6)))``. Short convolution adds |
| 9 | ``silu(sum_k weight[k] * history[t + k * dilation])`` to the gated output and |
| 10 | updates its persistent history. |
| 11 | """ |
| 12 | |
| 13 | from typing import Literal |
| 14 | |
| 15 | import torch |
| 16 | |
| 17 | from vllm.platforms import current_platform |
| 18 | from vllm.triton_utils import tl, triton |
| 19 | from vllm.utils.torch_utils import direct_register_custom_op |
| 20 | from vllm.v1.attention.backends.utils import NULL_BLOCK_ID |
| 21 | |
| 22 | |
| 23 | # N-gram ID generation |
| 24 | @triton.jit(do_not_specialize=["num_tokens", "num_reqs", "binary_search_iters"]) |
| 25 | def _ple_ngram_ids_kernel( |
| 26 | input_ids_ptr, |
| 27 | qsl_ptr, |
| 28 | ctx_ptr, |
| 29 | multipliers_ptr, |
| 30 | sizes_ptr, |
| 31 | offsets_ptr, |
| 32 | out_ptr, |
| 33 | num_tokens, |
| 34 | num_reqs, |
| 35 | eos_token_id, |
| 36 | binary_search_iters, |
| 37 | NGRAM_CONTEXT_LEN: tl.constexpr, |
| 38 | HEADS_PER_NGRAM: tl.constexpr, |
| 39 | BLOCK_T: tl.constexpr, |
| 40 | launch_pdl: tl.constexpr, |
| 41 | ): |
| 42 | NGRAM_HEADS: tl.constexpr = NGRAM_CONTEXT_LEN * HEADS_PER_NGRAM |
| 43 | BLOCK_H: tl.constexpr = triton.next_power_of_2(NGRAM_HEADS) |
| 44 | pid = tl.program_id(0) |
| 45 | token_offsets = pid * BLOCK_T + tl.arange(0, BLOCK_T) |
| 46 | token_mask = token_offsets < num_tokens |
| 47 | |
| 48 | if launch_pdl: |
| 49 | tl.extra.cuda.gdc_wait() |
| 50 | |
| 51 | # A flattened token offset does not identify its request when query chunks |
| 52 | # have different lengths. Binary-search the request boundaries instead. |
| 53 | p_lo = tl.full([BLOCK_T], 1, tl.int32) |
| 54 | p_hi = tl.full([BLOCK_T], num_reqs + 1, tl.int32) |
| 55 | for _ in range(binary_search_iters): |
| 56 | mid = (p_lo + p_hi) // 2 |
| 57 | qmid = tl.load(qsl_ptr + mid, mask=token_mask & (mid <= num_reqs), other=0) |
| 58 | pred = qmid <= token_offsets |
| 59 | p_lo = tl.where(pred, mid + 1, p_lo) |
| 60 | p_hi = tl.where(pred, p_hi, mid) |
| 61 | req = tl.minimum(p_lo - 1, num_reqs - 1).to(tl.int64) |
| 62 | request_start = tl.load(qsl_ptr + req, mask=token_mask, other=0) |
| 63 | chunk_pos = token_offsets - request_start |
| 64 | |
| 65 | # Initialize every head with the current-token term. |
| 66 | current_token = tl.load(input_ids_ptr + token_offsets, mask=token_mask, other=0).to( |
| 67 | tl.int64 |
| 68 | ) |
| 69 | current_multiplier = tl.load(multipliers_ptr) |
| 70 | mixed = current_token[:, None] * current_multiplier |
| 71 | |
| 72 | g = tl.arange(0, BLOCK_H) |
| 73 | head_mask = g < NGRAM_HEADS |
| 74 | ngram_order = g // HEADS_PER_NGRAM + 2 |
| 75 | |
| 76 | # Walk predecessors from newest to oldest. At chunk boundaries they come |
| 77 | # from ngram_context; otherwise they come from this step's input_ids. |
| 78 | crossed = tl.zeros([BLOCK_T], tl.int1) |
| 79 | for shift in tl.static_range(1, NGRAM_CONTEXT_LEN + 1): |
| 80 | in_step = chunk_pos >= shift |
| 81 | ctx_col = NGRAM_CONTEXT_LEN - shift + chunk_pos |
| 82 | step_token = tl.load( |
| 83 | input_ids_ptr + token_offsets - shift, |
| 84 | mask=token_mask & in_step, |
| 85 | other=0, |
| 86 | ) |
| 87 | context_token = tl.load( |
| 88 | ctx_ptr + req * NGRAM_CONTEXT_LEN + ctx_col, |
| 89 | mask=token_mask & (~in_step), |
| 90 | other=0, |
| 91 | ) |
| 92 | candidate = tl.where(in_step, step_token, context_token).to(tl.int64) |
| 93 | # Older positions remain behind the first EOS boundary. |
| 94 | candidate = tl.where(crossed, eos_token_id, candidate) |
| 95 | crossed = crossed | (candidate == eos_token_id) |
| 96 | multiplier = tl.load(multipliers_ptr + shift) |
| 97 | term = candidate[:, None] * multiplier |
| 98 | mixed = mixed ^ tl.where((ngram_order > shift)[None, :], term, 0) |
| 99 | |
| 100 | # Map each hash into its embedding-table partition. |
| 101 | sizes = tl.load(sizes_ptr + g, mask=head_mask, other=1)[None, :] |
| 102 | head_offsets = tl.load(offsets_ptr + g, mask=head_mask, other=0)[None, :] |
| 103 | # Hash products may overflow int64; preserve torch.remainder semantics. |
| 104 | remainders = mixed % sizes |
| 105 | remainders = tl.where(remainders < 0, remainders + sizes, remainders) |
| 106 | ids = remainders + head_offsets |
| 107 | if launch_pdl: |
| 108 | tl.extra.cuda.gdc_launch_dependents() |
| 109 | tl.store( |
| 110 | out_ptr + token_offsets[:, None] * NGRAM_HEADS + g[None, :], |
| 111 | ids, |
| 112 | mask=token_mask[:, None] & head_mask[None, :], |
| 113 | ) |
| 114 | |
| 115 | |
| 116 | def _ple_ngram_ids( |
| 117 | input_ids: torch.Tensor, |
| 118 | query_start_loc: torch.Tensor, |
| 119 | ngram_context: torch.Tensor, |
| 120 | layer_multipliers: torch.Tensor, |
| 121 | ngram_heads_vocab_sizes: torch.Tensor, |
| 122 | ngram_heads_offsets: torch.Tensor, |
| 123 | output: torch.Tensor, |
| 124 | eos_token_id: int, |
| 125 | heads_per_ngram: int, |
| 126 | ) -> None: |
| 127 | input_ids = input_ids.reshape(-1) |
| 128 | num_tokens = input_ids.shape[0] |
| 129 | num_reqs = query_start_loc.numel() - 1 |
| 130 | ctx_len = ngram_context.shape[1] |
| 131 | BLOCK_T = 8 |
| 132 | launch_pdl = current_platform.is_arch_support_pdl() |
| 133 | _ple_ngram_ids_kernel[(triton.cdiv(num_tokens, BLOCK_T),)]( |
| 134 | input_ids, |
| 135 | query_start_loc, |
| 136 | ngram_context, |
| 137 | layer_multipliers, |
| 138 | ngram_heads_vocab_sizes, |
| 139 | ngram_heads_offsets, |
| 140 | output, |
| 141 | num_tokens, |
| 142 | num_reqs, |
| 143 | eos_token_id, |
| 144 | binary_search_iters=num_reqs.bit_length(), |
| 145 | NGRAM_CONTEXT_LEN=ctx_len, |
| 146 | HEADS_PER_NGRAM=heads_per_ngram, |
| 147 | BLOCK_T=BLOCK_T, |
| 148 | launch_pdl=launch_pdl, |
| 149 | num_warps=4, |
| 150 | ) |
| 151 | |
| 152 | |
| 153 | def ple_ngram_ids( |
| 154 | input_ids: torch.Tensor, |
| 155 | query_start_loc: torch.Tensor, |
| 156 | ngram_context: torch.Tensor, |
| 157 | layer_multipliers: torch.Tensor, |
| 158 | ngram_heads_vocab_sizes: torch.Tensor, |
| 159 | ngram_heads_offsets: torch.Tensor, |
| 160 | eos_token_id: int, |
| 161 | heads_per_ngram: int, |
| 162 | output: torch.Tensor | None = None, |
| 163 | ) -> torch.Tensor: |
| 164 | if output is None: |
| 165 | output = torch.empty( |
| 166 | (input_ids.numel(), ngram_context.shape[1] * heads_per_ngram), |
| 167 | dtype=torch.int64, |
| 168 | device=input_ids.device, |
| 169 | ) |
| 170 | _ple_ngram_ids( |
| 171 | input_ids, |
| 172 | query_start_loc, |
| 173 | ngram_context, |
| 174 | layer_multipliers, |
| 175 | ngram_heads_vocab_sizes, |
| 176 | ngram_heads_offsets, |
| 177 | output, |
| 178 | eos_token_id, |
| 179 | heads_per_ngram, |
| 180 | ) |
| 181 | return output |
| 182 | |
| 183 | |
| 184 | # Gating and normalization |
| 185 | @triton.jit |
| 186 | def _ple_gate_kernel( |
| 187 | key_ptr, # [T, HC*H] with row stride key_rs |
| 188 | value_ptr, # [T, H] with row stride value_rs |
| 189 | hidden_ptr, # [T, HC*H] |
| 190 | nk_ptr, # [HC*H] norm_key weight |
| 191 | nq_ptr, # [HC*H] norm_query weight |
| 192 | ncw_ptr, # [HC*H] norm_conv weight |
| 193 | gated_ptr, # [T, HC*H] out |
| 194 | normed_ptr, # [T, HC*H] out |
| 195 | key_rs, |
| 196 | value_rs, |
| 197 | eps, |
| 198 | H: tl.constexpr, |
| 199 | HC: tl.constexpr, |
| 200 | BLOCK_H: tl.constexpr, |
| 201 | launch_pdl: tl.constexpr, |
| 202 | ): |
| 203 | t = tl.program_id(0) |
| 204 | s = tl.program_id(1) |
| 205 | lanes = tl.arange(0, BLOCK_H) |
| 206 | mask = lanes < H |
| 207 | offs = s * H + lanes |
| 208 | dtype: tl.constexpr = key_ptr.dtype.element_ty |
| 209 | |
| 210 | if launch_pdl: |
| 211 | tl.extra.cuda.gdc_wait() |
| 212 | |
| 213 | k = tl.load(key_ptr + t * key_rs + offs, mask=mask, other=0.0).to(tl.float32) |
| 214 | q = tl.load(hidden_ptr + t * HC * H + offs, mask=mask, other=0.0).to(tl.float32) |
| 215 | nk = tl.load(nk_ptr + offs, mask=mask, other=0.0).to(tl.float32) |
| 216 | nq = tl.load(nq_ptr + offs, mask=mask, other=0.0).to(tl.float32) |
| 217 | |
| 218 | # Match eager materialization at each intermediate tensor boundary. |
| 219 | k_n = (k * tl.rsqrt(tl.sum(k * k) / H + eps) * (1.0 + nk)).to(dtype) |
| 220 | q_n = (q * tl.rsqrt(tl.sum(q * q) / H + eps) * (1.0 + nq)).to(dtype) |
| 221 | products = (k_n.to(tl.float32) * q_n.to(tl.float32)).to(dtype) |
| 222 | dot = tl.sum(products.to(tl.float32)).to(dtype).to(tl.float32) |
| 223 | d = dot / tl.sqrt(float(H)) |
| 224 | d = d.to(dtype).to(tl.float32) |
| 225 | sign = tl.where(d < 0, -1.0, 0.0) |
| 226 | sign = tl.where(d > 0, 1.0, sign) |
| 227 | magnitude = tl.sqrt(tl.maximum(tl.abs(d), 1e-6)).to(dtype) |
| 228 | g = tl.sigmoid(sign * magnitude) |
| 229 | g = g.to(dtype).to(tl.float32) |
| 230 | |
| 231 | v = tl.load(value_ptr + t * value_rs + lanes, mask=mask, other=0.0).to(tl.float32) |
| 232 | gated = (g * v).to(dtype) |
| 233 | gf = gated.to(tl.float32) |
| 234 | ncw = tl.load(ncw_ptr + offs, mask=mask, other=0.0).to(tl.float32) |
| 235 | normed = gf * tl.rsqrt(tl.sum(gf * gf) / H + eps) * (1.0 + ncw) |
| 236 | |
| 237 | if launch_pdl: |
| 238 | tl.extra.cuda.gdc_launch_dependents() |
| 239 | tl.store(gated_ptr + t * HC * H + offs, gated, mask=mask) |
| 240 | tl.store(normed_ptr + t * HC * H + offs, normed, mask=mask) |
| 241 | |
| 242 | |
| 243 | def _ple_gate( |
| 244 | key: torch.Tensor, |
| 245 | value: torch.Tensor, |
| 246 | hidden: torch.Tensor, |
| 247 | norm_key_w: torch.Tensor, |
| 248 | norm_query_w: torch.Tensor, |
| 249 | norm_conv_w: torch.Tensor, |
| 250 | eps: float, |
| 251 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 252 | num_tokens = hidden.shape[0] |
| 253 | h = value.shape[-1] |
| 254 | hc = hidden.shape[-1] // h |
| 255 | assert key.stride(1) == 1 and value.stride(1) == 1 |
| 256 | assert hidden.is_contiguous() |
| 257 | if key.dtype != value.dtype or key.dtype != hidden.dtype: |
| 258 | raise ValueError("key, value, and hidden must have the same dtype") |
| 259 | if key.dtype not in (torch.bfloat16, torch.float16): |
| 260 | raise ValueError("PLE gate supports BF16 and FP16 inputs") |
| 261 | gated = torch.empty_like(hidden) |
| 262 | normed = torch.empty_like(hidden) |
| 263 | _ple_gate_kernel[(num_tokens, hc)]( |
| 264 | key, |
| 265 | value, |
| 266 | hidden, |
| 267 | norm_key_w, |
| 268 | norm_query_w, |
| 269 | norm_conv_w, |
| 270 | gated, |
| 271 | normed, |
| 272 | key.stride(0), |
| 273 | value.stride(0), |
| 274 | eps, |
| 275 | H=h, |
| 276 | HC=hc, |
| 277 | BLOCK_H=triton.next_power_of_2(h), |
| 278 | num_warps=4, |
| 279 | launch_pdl=current_platform.is_arch_support_pdl(), |
| 280 | ) |
| 281 | return gated, normed |
| 282 | |
| 283 | |
| 284 | def _ple_gate_fake( |
| 285 | key: torch.Tensor, |
| 286 | value: torch.Tensor, |
| 287 | hidden: torch.Tensor, |
| 288 | norm_key_w: torch.Tensor, |
| 289 | norm_query_w: torch.Tensor, |
| 290 | norm_conv_w: torch.Tensor, |
| 291 | eps: float, |
| 292 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 293 | return torch.empty_like(hidden), torch.empty_like(hidden) |
| 294 | |
| 295 | |
| 296 | direct_register_custom_op( |
| 297 | op_name="qwen4_exp_ple_gate", |
| 298 | op_func=_ple_gate, |
| 299 | mutates_args=[], |
| 300 | fake_impl=_ple_gate_fake, |
| 301 | ) |
| 302 | |
| 303 | |
| 304 | def ple_gate( |
| 305 | key: torch.Tensor, |
| 306 | value: torch.Tensor, |
| 307 | hidden: torch.Tensor, |
| 308 | norm_key_w: torch.Tensor, |
| 309 | norm_query_w: torch.Tensor, |
| 310 | norm_conv_w: torch.Tensor, |
| 311 | eps: float, |
| 312 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 313 | return torch.ops.vllm.qwen4_exp_ple_gate( |
| 314 | key, value, hidden, norm_key_w, norm_query_w, norm_conv_w, eps |
| 315 | ) |
| 316 | |
| 317 | |
| 318 | # Dilated short convolution |
| 319 | @triton.jit(do_not_specialize=["num_reqs", "bs_iters", "has_token_map"]) |
| 320 | def _ple_conv_kernel( |
| 321 | x_ptr, |
| 322 | state_ptr, |
| 323 | w_ptr, |
| 324 | residual_ptr, |
| 325 | outer_residual_ptr, |
| 326 | state_idx_ptr, |
| 327 | qsl_ptr, |
| 328 | num_acc_ptr, |
| 329 | has_init_ptr, |
| 330 | token_idx_ptr, |
| 331 | has_token_map, |
| 332 | num_reqs, |
| 333 | bs_iters, |
| 334 | state_idx_stride, |
| 335 | state_bs, |
| 336 | state_ws, |
| 337 | state_cs, |
| 338 | C: tl.constexpr, |
| 339 | BLOCK_C: tl.constexpr, |
| 340 | STATE_LEN: tl.constexpr, |
| 341 | DILATION: tl.constexpr, |
| 342 | KERNEL_SIZE: tl.constexpr, |
| 343 | SPEC_QUERY_LEN: tl.constexpr, |
| 344 | MODE: tl.constexpr, |
| 345 | HAS_INIT: tl.constexpr, |
| 346 | NULL_STATE_ID: tl.constexpr, |
| 347 | launch_pdl: tl.constexpr, |
| 348 | ): |
| 349 | t = tl.program_id(0) |
| 350 | pid_c = tl.program_id(1) |
| 351 | c_offs = pid_c * BLOCK_C + tl.arange(0, BLOCK_C) |
| 352 | c_mask = c_offs < C |
| 353 | |
| 354 | if launch_pdl: |
| 355 | tl.extra.cuda.gdc_wait() |
| 356 | |
| 357 | # Mode-specific launches use logical row indices. Mixed batches map them |
| 358 | # back to the original token order without gathering inputs or residuals. |
| 359 | output_t = tl.load(token_idx_ptr + t, mask=has_token_map, other=t).to(tl.int64) |
| 360 | |
| 361 | if MODE == "decode": |
| 362 | r = t |
| 363 | q_start = t |
| 364 | j = tl.full([], 0, tl.int32) |
| 365 | slot_off = tl.full([], 0, tl.int32) |
| 366 | in_range = True |
| 367 | else: |
| 368 | # Locate the request containing this token. |
| 369 | lo = tl.full([], 1, tl.int32) |
| 370 | hi = tl.full([], num_reqs + 1, tl.int32) |
| 371 | for _ in range(bs_iters): |
| 372 | mid = (lo + hi) // 2 |
| 373 | qmid = tl.load(qsl_ptr + mid, mask=mid <= num_reqs, other=0) |
| 374 | pred = qmid <= t |
| 375 | lo = tl.where(pred, mid + 1, lo) |
| 376 | hi = tl.where(pred, hi, mid) |
| 377 | r = tl.minimum(lo - 1, num_reqs - 1) |
| 378 | q_start = tl.load(qsl_ptr + r) |
| 379 | j = (t - q_start).to(tl.int32) |
| 380 | total_real = tl.load(qsl_ptr + num_reqs) |
| 381 | in_range = t < total_real |
| 382 | if MODE == "spec": |
| 383 | num_acc = tl.load(num_acc_ptr + r) |
| 384 | # Roll back draft state to the last accepted token. |
| 385 | slot_off = tl.minimum(tl.maximum(num_acc - 1, 0), SPEC_QUERY_LEN - 1).to( |
| 386 | tl.int32 |
| 387 | ) |
| 388 | else: |
| 389 | slot_off = tl.full([], 0, tl.int32) |
| 390 | |
| 391 | sid = tl.load(state_idx_ptr + r * state_idx_stride).to(tl.int64) |
| 392 | state_ok = sid != NULL_STATE_ID |
| 393 | sid_safe = tl.where(state_ok, sid, 0) |
| 394 | if HAS_INIT: |
| 395 | has_init = tl.load(has_init_ptr + r, mask=state_ok, other=0) != 0 |
| 396 | else: |
| 397 | has_init = state_ok |
| 398 | if MODE == "spec": |
| 399 | # A new spec chunk can produce output from its own tokens without a |
| 400 | # prior state; null slots in other modes denote inactive rows. |
| 401 | read_state = state_ok |
| 402 | out_ok = in_range |
| 403 | else: |
| 404 | read_state = state_ok & has_init |
| 405 | out_ok = in_range & state_ok |
| 406 | |
| 407 | # state[sid, w, c] may be a strided view into a shared pool. |
| 408 | base_state = state_ptr + sid_safe * state_bs |
| 409 | acc = tl.zeros([BLOCK_C], tl.float32) |
| 410 | for k in tl.static_range(0, KERNEL_SIZE): |
| 411 | h = j + DILATION * k |
| 412 | from_state = h <= STATE_LEN - 1 |
| 413 | state_tap = tl.load( |
| 414 | base_state + (h + slot_off) * state_ws + c_offs * state_cs, |
| 415 | mask=c_mask & read_state & from_state, |
| 416 | other=0.0, |
| 417 | ) |
| 418 | input_t = q_start + h - STATE_LEN |
| 419 | input_t = tl.load( |
| 420 | token_idx_ptr + input_t, |
| 421 | mask=has_token_map & out_ok & (~from_state), |
| 422 | other=input_t, |
| 423 | ).to(tl.int64) |
| 424 | input_tap = tl.load( |
| 425 | x_ptr + input_t * C + c_offs, |
| 426 | mask=c_mask & out_ok & (~from_state), |
| 427 | other=0.0, |
| 428 | ) |
| 429 | tap = tl.where(from_state, state_tap, input_tap).to(tl.float32) |
| 430 | weight = tl.load( |
| 431 | w_ptr + c_offs * KERNEL_SIZE + k, |
| 432 | mask=c_mask, |
| 433 | other=0.0, |
| 434 | ).to(tl.float32) |
| 435 | acc += weight * tap |
| 436 | |
| 437 | # F.conv1d materializes its output dtype before SiLU. |
| 438 | conv = acc.to(residual_ptr.dtype.element_ty).to(tl.float32) |
| 439 | y = conv * tl.sigmoid(conv) |
| 440 | conv_output = tl.where(out_ok, y, 0.0).to(residual_ptr.dtype.element_ty) |
| 441 | residual = tl.load( |
| 442 | residual_ptr + output_t * C + c_offs, |
| 443 | mask=c_mask, |
| 444 | other=0.0, |
| 445 | ) |
| 446 | # Preserve the original eager operation boundaries: short convolution is |
| 447 | # first accumulated into the BF16/FP16 PLE output, then the outer residual |
| 448 | # is added to that rounded value. |
| 449 | ple_output = (residual + conv_output).to(residual_ptr.dtype.element_ty) |
| 450 | outer_residual = tl.load( |
| 451 | outer_residual_ptr + output_t * C + c_offs, |
| 452 | mask=c_mask, |
| 453 | other=0.0, |
| 454 | ) |
| 455 | ple_output = outer_residual.to(tl.float32) + ple_output.to(tl.float32) |
| 456 | if launch_pdl: |
| 457 | tl.extra.cuda.gdc_launch_dependents() |
| 458 | tl.store( |
| 459 | residual_ptr + output_t * C + c_offs, |
| 460 | ple_output, |
| 461 | mask=c_mask, |
| 462 | ) |
| 463 | |
| 464 | if MODE == "decode": |
| 465 | # Decode has one program per request and channel block, so no peer can |
| 466 | # still be reading this state slice when it is updated. |
| 467 | decode_input = tl.load( |
| 468 | x_ptr + output_t * C + c_offs, |
| 469 | mask=c_mask & state_ok, |
| 470 | other=0.0, |
| 471 | ) |
| 472 | for i in tl.static_range(0, STATE_LEN): |
| 473 | if i < STATE_LEN - 1: |
| 474 | next_state = tl.load( |
| 475 | base_state + (i + 1) * state_ws + c_offs * state_cs, |
| 476 | mask=c_mask & state_ok & has_init, |
| 477 | other=0.0, |
| 478 | ) |
| 479 | else: |
| 480 | next_state = decode_input |
| 481 | tl.store( |
| 482 | base_state + i * state_ws + c_offs * state_cs, |
| 483 | next_state, |
| 484 | mask=c_mask & state_ok, |
| 485 | ) |
| 486 | |
| 487 | |
| 488 | # Prefill and spec write back separately so every token reads the old state. |
| 489 | @triton.jit(do_not_specialize=["has_token_map"]) |
| 490 | def _ple_conv_writeback_kernel( |
| 491 | x_ptr, |
| 492 | state_ptr, |
| 493 | state_idx_ptr, |
| 494 | qsl_ptr, |
| 495 | num_acc_ptr, |
| 496 | has_init_ptr, |
| 497 | token_idx_ptr, |
| 498 | has_token_map, |
| 499 | state_idx_stride, |
| 500 | state_bs, |
| 501 | state_ws, |
| 502 | state_cs, |
| 503 | C: tl.constexpr, |
| 504 | BLOCK_C: tl.constexpr, |
| 505 | STATE_LEN: tl.constexpr, |
| 506 | SPEC_QUERY_LEN: tl.constexpr, |
| 507 | STATE_WIDTH: tl.constexpr, |
| 508 | MODE: tl.constexpr, |
| 509 | HAS_INIT: tl.constexpr, |
| 510 | NULL_STATE_ID: tl.constexpr, |
| 511 | ): |
| 512 | r = tl.program_id(0) |
| 513 | pid_c = tl.program_id(1) |
| 514 | c_offs = pid_c * BLOCK_C + tl.arange(0, BLOCK_C) |
| 515 | c_mask = c_offs < C |
| 516 | |
| 517 | sid = tl.load(state_idx_ptr + r * state_idx_stride).to(tl.int64) |
| 518 | state_ok = sid != NULL_STATE_ID |
| 519 | if not state_ok: |
| 520 | return |
| 521 | |
| 522 | q_start = tl.load(qsl_ptr + r) |
| 523 | q_end = tl.load(qsl_ptr + r + 1) |
| 524 | qlen = (q_end - q_start).to(tl.int32) |
| 525 | if MODE == "spec": |
| 526 | num_acc = tl.load(num_acc_ptr + r) |
| 527 | # Use the same rollback window as the output kernel. |
| 528 | slot_off = tl.minimum(tl.maximum(num_acc - 1, 0), SPEC_QUERY_LEN - 1).to( |
| 529 | tl.int32 |
| 530 | ) |
| 531 | shift = 1 |
| 532 | else: |
| 533 | slot_off = tl.full([], 0, tl.int32) |
| 534 | shift = qlen |
| 535 | if qlen <= 0: |
| 536 | return |
| 537 | |
| 538 | has_init = tl.load(has_init_ptr + r) != 0 if HAS_INIT else True |
| 539 | src_ok = state_ok if MODE == "spec" else state_ok & has_init |
| 540 | |
| 541 | WRITE_W: tl.constexpr = STATE_WIDTH if MODE == "spec" else STATE_LEN |
| 542 | base_state = state_ptr + sid * state_bs |
| 543 | for i in tl.static_range(0, WRITE_W): |
| 544 | m = shift + i |
| 545 | from_state = m <= STATE_LEN - 1 |
| 546 | # Short or graph-padded spec chunks preserve the unused state tail. |
| 547 | do_write = i < STATE_LEN + qlen - 1 if MODE == "spec" else True |
| 548 | state_value = tl.load( |
| 549 | base_state + (slot_off + m) * state_ws + c_offs * state_cs, |
| 550 | mask=c_mask & from_state & src_ok, |
| 551 | other=0.0, |
| 552 | ) |
| 553 | input_t = q_start + m - STATE_LEN |
| 554 | input_t = tl.load( |
| 555 | token_idx_ptr + input_t, |
| 556 | mask=has_token_map & (~from_state) & do_write, |
| 557 | other=input_t, |
| 558 | ).to(tl.int64) |
| 559 | input_value = tl.load( |
| 560 | x_ptr + input_t * C + c_offs, |
| 561 | mask=c_mask & (~from_state) & do_write, |
| 562 | other=0.0, |
| 563 | ) |
| 564 | value = tl.where(from_state, state_value, input_value) |
| 565 | tl.store( |
| 566 | base_state + i * state_ws + c_offs * state_cs, |
| 567 | value, |
| 568 | mask=c_mask & do_write, |
| 569 | ) |
| 570 | |
| 571 | |
| 572 | def ple_conv( |
| 573 | inputs: torch.Tensor, |
| 574 | residual: torch.Tensor, |
| 575 | conv_state: torch.Tensor, |
| 576 | conv_weights: torch.Tensor, |
| 577 | state_indices: torch.Tensor, |
| 578 | outer_residual: torch.Tensor, |
| 579 | *, |
| 580 | mode: Literal["decode", "spec", "prefill"], |
| 581 | dilation: int, |
| 582 | query_start_loc: torch.Tensor | None = None, |
| 583 | num_accepted_tokens: torch.Tensor | None = None, |
| 584 | has_initial_states: torch.Tensor | None = None, |
| 585 | spec_query_len: int = 1, |
| 586 | token_indices: torch.Tensor | None = None, |
| 587 | ) -> None: |
| 588 | """Add short convolution and the outer residual; update state.""" |
| 589 | BLOCK_C = 512 |
| 590 | kernel_spec_query_len = spec_query_len if mode == "spec" else 1 |
| 591 | T, C = inputs.shape |
| 592 | K = conv_weights.shape[1] |
| 593 | state_len = (K - 1) * dilation |
| 594 | state_width = state_len + kernel_spec_query_len - 1 |
| 595 | if token_indices is not None: |
| 596 | T = token_indices.numel() |
| 597 | if conv_state.shape[1] != C or conv_state.shape[2] < state_width: |
| 598 | raise ValueError( |
| 599 | "conv_state must have shape [slots, channels, window], with " |
| 600 | f"channels={C} and window >= {state_width}" |
| 601 | ) |
| 602 | state_bs, state_cs, state_ws = conv_state.stride() |
| 603 | |
| 604 | if mode == "decode": |
| 605 | num_reqs = T |
| 606 | binary_search_iters = 1 |
| 607 | has_initial_states_arg = has_initial_states is not None |
| 608 | elif mode == "spec": |
| 609 | if query_start_loc is None or num_accepted_tokens is None: |
| 610 | raise ValueError( |
| 611 | "query_start_loc and num_accepted_tokens are required for spec decode" |
| 612 | ) |
| 613 | num_reqs = state_indices.numel() |
| 614 | binary_search_iters = max(num_reqs, 1).bit_length() |
| 615 | has_initial_states_arg = False |
| 616 | elif mode == "prefill": |
| 617 | if query_start_loc is None or has_initial_states is None: |
| 618 | raise ValueError( |
| 619 | "query_start_loc and has_initial_states are required for prefill" |
| 620 | ) |
| 621 | num_reqs = state_indices.numel() |
| 622 | binary_search_iters = max(num_reqs, 1).bit_length() |
| 623 | has_initial_states_arg = True |
| 624 | else: |
| 625 | raise ValueError(f"Unsupported short-conv mode: {mode}") |
| 626 | |
| 627 | num_warps = 4 if mode == "prefill" else 8 |
| 628 | launch_pdl = current_platform.is_arch_support_pdl() |
| 629 | # Pure-prefill indices can be a strided block-table column view. |
| 630 | state_idx_stride = state_indices.stride(0) |
| 631 | |
| 632 | # Constexpr flags eliminate accesses to optional None arguments. Without a |
| 633 | # token map, state_indices is an unused but device-resident placeholder. |
| 634 | _ple_conv_kernel[(T, triton.cdiv(C, BLOCK_C))]( |
| 635 | inputs, |
| 636 | conv_state, |
| 637 | conv_weights, |
| 638 | residual, |
| 639 | outer_residual, |
| 640 | state_indices, |
| 641 | query_start_loc, |
| 642 | num_accepted_tokens, |
| 643 | has_initial_states, |
| 644 | token_indices if token_indices is not None else state_indices, |
| 645 | token_indices is not None, |
| 646 | num_reqs, |
| 647 | binary_search_iters, |
| 648 | state_idx_stride, |
| 649 | state_bs, |
| 650 | state_ws, |
| 651 | state_cs, |
| 652 | C=C, |
| 653 | BLOCK_C=BLOCK_C, |
| 654 | STATE_LEN=state_len, |
| 655 | DILATION=dilation, |
| 656 | KERNEL_SIZE=K, |
| 657 | SPEC_QUERY_LEN=kernel_spec_query_len, |
| 658 | MODE=mode, |
| 659 | HAS_INIT=has_initial_states_arg, |
| 660 | NULL_STATE_ID=NULL_BLOCK_ID, |
| 661 | launch_pdl=launch_pdl, |
| 662 | num_warps=num_warps, |
| 663 | ) |
| 664 | # conv state update is fused with the kernel above for decode |
| 665 | if mode != "decode": |
| 666 | _ple_conv_writeback_kernel[(num_reqs, triton.cdiv(C, BLOCK_C))]( |
| 667 | inputs, |
| 668 | conv_state, |
| 669 | state_indices, |
| 670 | query_start_loc, |
| 671 | num_accepted_tokens, |
| 672 | has_initial_states, |
| 673 | token_indices if token_indices is not None else state_indices, |
| 674 | token_indices is not None, |
| 675 | state_idx_stride, |
| 676 | state_bs, |
| 677 | state_ws, |
| 678 | state_cs, |
| 679 | C=C, |
| 680 | BLOCK_C=BLOCK_C, |
| 681 | STATE_LEN=state_len, |
| 682 | SPEC_QUERY_LEN=kernel_spec_query_len, |
| 683 | STATE_WIDTH=state_width, |
| 684 | MODE=mode, |
| 685 | HAS_INIT=has_initial_states_arg, |
| 686 | NULL_STATE_ID=NULL_BLOCK_ID, |
| 687 | num_warps=num_warps, |
| 688 | ) |