| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """CuTe DSL sparse-attention compressor for DeepSeek V4. |
| 4 | |
| 5 | The public wrappers provide the C4 fused and C128 split kernels. |
| 6 | """ |
| 7 | |
| 8 | from __future__ import annotations |
| 9 | |
| 10 | from dataclasses import dataclass |
| 11 | from typing import Any |
| 12 | |
| 13 | import cutlass |
| 14 | import cutlass.cute as cute |
| 15 | import torch |
| 16 | from cuda.bindings.driver import CUstream |
| 17 | from cutlass import Float32, Int32, Int64, Uint8, Uint16, Uint32, const_expr |
| 18 | from cutlass._mlir.dialects import llvm |
| 19 | from cutlass.cutlass_dsl import T, dsl_user_op |
| 20 | from quack.compile_utils import make_fake_tensor |
| 21 | |
| 22 | from vllm.cute_utils import torch_to_cute_dtype |
| 23 | from vllm.model_executor.warmup.jit_warmup import kernel_launcher |
| 24 | from vllm.model_executor.warmup.jit_warmup_cutedsl_helper import ( |
| 25 | CuTeDSLLaunchSpec, |
| 26 | VllmCuTeDSLJitKernel, |
| 27 | ) |
| 28 | from vllm.utils.math_utils import round_up |
| 29 | |
| 30 | |
| 31 | @dsl_user_op |
| 32 | def _recast_val(x, dtype, *, loc=None, ip=None): |
| 33 | return dtype(llvm.bitcast(dtype.mlir_type, x.ir_value(loc=loc, ip=ip))) |
| 34 | |
| 35 | |
| 36 | @dsl_user_op |
| 37 | def _fp32x2_to_bf16x2(a: Float32, b: Float32, *, loc=None, ip=None) -> Uint32: |
| 38 | out = llvm.inline_asm( |
| 39 | T.i32(), |
| 40 | [a.ir_value(loc=loc, ip=ip), b.ir_value(loc=loc, ip=ip)], |
| 41 | "cvt.rn.bf16x2.f32 $0, $2, $1;", |
| 42 | "=r,f,f", |
| 43 | has_side_effects=False, |
| 44 | is_align_stack=False, |
| 45 | ) |
| 46 | return Uint32(out) |
| 47 | |
| 48 | |
| 49 | @dsl_user_op |
| 50 | def _bf16x2_to_fp32(data: Uint32, *, loc=None, ip=None) -> tuple[Float32, Float32]: |
| 51 | out = llvm.inline_asm( |
| 52 | llvm.StructType.get_literal([T.f32(), T.f32()]), |
| 53 | [data.ir_value(loc=loc, ip=ip)], |
| 54 | "shl.b32 $0, $2, 16;\n\tand.b32 $1, $2, 0xFFFF0000;\n", |
| 55 | "=f,=f,r", |
| 56 | has_side_effects=False, |
| 57 | is_align_stack=False, |
| 58 | ) |
| 59 | return ( |
| 60 | Float32(llvm.extractvalue(T.f32(), out, [0], loc=loc, ip=ip)), |
| 61 | Float32(llvm.extractvalue(T.f32(), out, [1], loc=loc, ip=ip)), |
| 62 | ) |
| 63 | |
| 64 | |
| 65 | @dsl_user_op |
| 66 | def _fp32x2_to_fp8e4m3x2(a: Float32, b: Float32, *, loc=None, ip=None) -> Uint16: |
| 67 | out = llvm.inline_asm( |
| 68 | T.i16(), |
| 69 | [a.ir_value(loc=loc, ip=ip), b.ir_value(loc=loc, ip=ip)], |
| 70 | "cvt.rn.satfinite.e4m3x2.f32 $0, $2, $1;", |
| 71 | "=h,f,f", |
| 72 | has_side_effects=False, |
| 73 | is_align_stack=False, |
| 74 | ) |
| 75 | return Uint16(out) |
| 76 | |
| 77 | |
| 78 | class SparseAttnCompressNormRopeStoreC4Kernel( |
| 79 | VllmCuTeDSLJitKernel["SparseAttnCompressNormRopeStoreC4Kernel.CompileKey"] |
| 80 | ): |
| 81 | min_scale = 1.0e-4 |
| 82 | rcp_ln2 = 1.4426950408889634 |
| 83 | |
| 84 | @dataclass(frozen=True) |
| 85 | class CompileKey: |
| 86 | head_size: int |
| 87 | state_width: int |
| 88 | rope_head_dim: int |
| 89 | fp8_max: float |
| 90 | quant_block: int |
| 91 | token_stride: int |
| 92 | scale_dim: int |
| 93 | compress_ratio: int |
| 94 | overlap: bool |
| 95 | norm_weight_dtype: type[cutlass.Numeric] |
| 96 | |
| 97 | @staticmethod |
| 98 | def kernel(compile_key: CompileKey) -> Any: |
| 99 | min_scale = SparseAttnCompressNormRopeStoreC4Kernel.min_scale |
| 100 | rcp_ln2 = SparseAttnCompressNormRopeStoreC4Kernel.rcp_ln2 |
| 101 | head_dim = compile_key.head_size |
| 102 | state_width = compile_key.state_width |
| 103 | rope_dim = compile_key.rope_head_dim |
| 104 | nope_dim = head_dim - rope_dim |
| 105 | fp8_max = compile_key.fp8_max |
| 106 | quant_block = compile_key.quant_block |
| 107 | token_stride = compile_key.token_stride |
| 108 | scale_dim = compile_key.scale_dim |
| 109 | elems_per_lane = 8 |
| 110 | copy_elems = 4 |
| 111 | copy_chunks = elems_per_lane // copy_elems |
| 112 | lanes_per_group = quant_block // elems_per_lane |
| 113 | groups_per_warp = 32 // lanes_per_group |
| 114 | scale_reduce_steps = lanes_per_group.bit_length() - 1 |
| 115 | scale_reduce_offset = lanes_per_group // 2 |
| 116 | num_warps = (head_dim // quant_block) // groups_per_warp |
| 117 | nope_blocks = nope_dim // quant_block |
| 118 | tb_size = num_warps * 32 |
| 119 | compress_ratio = compile_key.compress_ratio |
| 120 | overlap = compile_key.overlap |
| 121 | window = (1 + int(overlap)) * compress_ratio |
| 122 | |
| 123 | @cute.kernel |
| 124 | def device_kernel( |
| 125 | state_cache: cute.Tensor, |
| 126 | token_to_req_indices: cute.Tensor, |
| 127 | positions: cute.Tensor, |
| 128 | slot_mapping: cute.Tensor, |
| 129 | block_table: cute.Tensor, |
| 130 | block_size: Int64, |
| 131 | rms_norm_weight: cute.Tensor, |
| 132 | rms_norm_eps: Float32, |
| 133 | cos_sin_cache: cute.Tensor, |
| 134 | k_cache: cute.Tensor, |
| 135 | kv_slot_mapping: cute.Tensor, |
| 136 | kv_cache_block_size: Int64, |
| 137 | ): |
| 138 | token_idx, _, _ = cute.arch.block_idx() |
| 139 | tid, _, _ = cute.arch.thread_idx() |
| 140 | warp_id = cute.arch.make_warp_uniform(tid // 32) |
| 141 | lane_id = tid % 32 |
| 142 | group_lane = lane_id % lanes_per_group |
| 143 | group_idx = warp_id * groups_per_warp + lane_id // lanes_per_group |
| 144 | elem_base = group_idx * quant_block + group_lane * elems_per_lane |
| 145 | |
| 146 | slot_id = slot_mapping[token_idx] |
| 147 | has_position = token_idx < positions.shape[0] |
| 148 | position = Int64(0) |
| 149 | if has_position: |
| 150 | position = positions[token_idx] |
| 151 | boundary = has_position and ( |
| 152 | (position + Int64(1)) % Int64(compress_ratio) == Int64(0) |
| 153 | ) |
| 154 | has_req_idx = token_idx < token_to_req_indices.shape[0] |
| 155 | has_kv_slot_idx = token_idx < kv_slot_mapping.shape[0] |
| 156 | kv_slot_idx = Int64(-1) |
| 157 | if has_kv_slot_idx: |
| 158 | kv_slot_idx = kv_slot_mapping[token_idx] |
| 159 | active = ( |
| 160 | slot_id >= Int64(0) |
| 161 | and has_req_idx |
| 162 | and boundary |
| 163 | and kv_slot_idx >= Int64(0) |
| 164 | ) |
| 165 | |
| 166 | if active: |
| 167 | req_idx = token_to_req_indices[token_idx] |
| 168 | start = position - Int64(window - 1) |
| 169 | |
| 170 | smem = cutlass.utils.SmemAllocator() |
| 171 | s_block_numbers = smem.allocate_tensor( |
| 172 | Int32, cute.make_layout((window,)), byte_alignment=4 |
| 173 | ) |
| 174 | partial_sums = smem.allocate_tensor( |
| 175 | Float32, cute.make_layout((num_warps,)), byte_alignment=4 |
| 176 | ) |
| 177 | rrms_shared = smem.allocate_tensor( |
| 178 | Float32, cute.make_layout((1,)), byte_alignment=4 |
| 179 | ) |
| 180 | |
| 181 | for row in cutlass.range_constexpr(window): |
| 182 | pos = start + Int64(row) |
| 183 | if tid == row: |
| 184 | block_number_i32 = Int32(0) |
| 185 | if pos >= Int64(0): |
| 186 | block_index = pos // block_size |
| 187 | block_number_i32 = block_table[req_idx, block_index] |
| 188 | s_block_numbers[row] = block_number_i32 |
| 189 | cute.arch.sync_threads() |
| 190 | |
| 191 | local_max = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 192 | local_sum = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 193 | local_product = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 194 | |
| 195 | for e in cutlass.range_constexpr(elems_per_lane): |
| 196 | local_max[e] = -Float32.inf |
| 197 | local_sum[e] = Float32(0.0) |
| 198 | local_product[e] = Float32(0.0) |
| 199 | |
| 200 | cp_f32x4 = cute.make_copy_atom( |
| 201 | cute.nvgpu.CopyUniversalOp(), Float32, num_bits_per_copy=128 |
| 202 | ) |
| 203 | copy_layout = cute.make_layout( |
| 204 | (copy_chunks, copy_elems), |
| 205 | stride=(copy_elems, 1), |
| 206 | ) |
| 207 | kv_vals = cute.make_rmem_tensor(copy_layout, Float32) |
| 208 | score_vals = cute.make_rmem_tensor(copy_layout, Float32) |
| 209 | |
| 210 | for row in cutlass.range_constexpr(window): |
| 211 | pos = start + Int64(row) |
| 212 | if pos >= Int64(0): |
| 213 | block_index = pos // block_size |
| 214 | block_offset = pos - block_index * block_size |
| 215 | block_number = s_block_numbers[row].to(Int64) |
| 216 | head_offset = Int64((row // compress_ratio) * head_dim) |
| 217 | row_tensor = state_cache[block_number, block_offset, None] |
| 218 | for chunk in cutlass.range_constexpr(copy_chunks): |
| 219 | copy_elem = const_expr(chunk * copy_elems) |
| 220 | col_tile = ( |
| 221 | head_offset + (elem_base + Int32(copy_elem)).to(Int64) |
| 222 | ) // Int64(copy_elems) |
| 223 | kv_src = cute.local_tile( |
| 224 | row_tensor, |
| 225 | tiler=(copy_elems,), |
| 226 | coord=(col_tile,), |
| 227 | ) |
| 228 | score_src = cute.local_tile( |
| 229 | row_tensor, |
| 230 | tiler=(copy_elems,), |
| 231 | coord=(col_tile + Int64(state_width // copy_elems),), |
| 232 | ) |
| 233 | cute.copy(cp_f32x4, kv_src, kv_vals[chunk, None]) |
| 234 | cute.copy(cp_f32x4, score_src, score_vals[chunk, None]) |
| 235 | |
| 236 | for e in cutlass.range_constexpr(elems_per_lane): |
| 237 | chunk = const_expr(e // copy_elems) |
| 238 | copy_elem = const_expr(e % copy_elems) |
| 239 | score = score_vals[chunk, copy_elem] |
| 240 | kv = kv_vals[chunk, copy_elem] |
| 241 | new_max = cute.arch.fmax(local_max[e], score) |
| 242 | old_scale = cute.math.exp2( |
| 243 | (local_max[e] - new_max) * Float32(rcp_ln2), |
| 244 | fastmath=True, |
| 245 | ) |
| 246 | new_scale = cute.math.exp2( |
| 247 | (score - new_max) * Float32(rcp_ln2), |
| 248 | fastmath=True, |
| 249 | ) |
| 250 | local_sum[e] = local_sum[e] * old_scale + new_scale |
| 251 | local_product[e] = ( |
| 252 | local_product[e] * old_scale + kv * new_scale |
| 253 | ) |
| 254 | local_max[e] = new_max |
| 255 | |
| 256 | x = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 257 | local_sumsq = Float32(0.0) |
| 258 | for e in cutlass.range_constexpr(elems_per_lane): |
| 259 | x[e] = local_product[e] / local_sum[e] |
| 260 | local_sumsq += x[e] * x[e] |
| 261 | |
| 262 | warp_sum = local_sumsq |
| 263 | for step in cutlass.range_constexpr(5): |
| 264 | offset = const_expr(16 >> step) |
| 265 | warp_sum += cute.arch.shuffle_sync_bfly(warp_sum, offset) |
| 266 | |
| 267 | if lane_id == 0: |
| 268 | partial_sums[warp_id] = warp_sum |
| 269 | cute.arch.sync_threads() |
| 270 | if tid == 0: |
| 271 | total = Float32(0.0) |
| 272 | for i in cutlass.range_constexpr(num_warps): |
| 273 | total += partial_sums[i] |
| 274 | rrms_shared[0] = cute.math.rsqrt( |
| 275 | total / Float32(head_dim) + rms_norm_eps, fastmath=True |
| 276 | ) |
| 277 | cute.arch.sync_threads() |
| 278 | |
| 279 | rrms = rrms_shared[0] |
| 280 | for e in cutlass.range_constexpr(elems_per_lane): |
| 281 | elem = elem_base + e |
| 282 | x[e] = x[e] * rrms * rms_norm_weight[elem].to(Float32) |
| 283 | |
| 284 | k_cache_u16 = cute.recast_tensor(k_cache, Uint16) |
| 285 | k_cache_u32 = cute.recast_tensor(k_cache, Uint32) |
| 286 | page = kv_slot_idx // kv_cache_block_size |
| 287 | kv_offset = kv_slot_idx - page * kv_cache_block_size |
| 288 | value_base = page * k_cache.stride[0] + kv_offset * Int64(token_stride) |
| 289 | scale_base = ( |
| 290 | page * k_cache.stride[0] |
| 291 | + kv_cache_block_size * Int64(token_stride) |
| 292 | + kv_offset * Int64(scale_dim) |
| 293 | ) |
| 294 | |
| 295 | if group_idx == nope_blocks: |
| 296 | compressed_pos = (position // Int64(compress_ratio)) * Int64( |
| 297 | compress_ratio |
| 298 | ) |
| 299 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 300 | elem = const_expr(pair * 2) |
| 301 | pair_idx = (elem_base - nope_dim) // 2 + Int32(pair) |
| 302 | cos_v = cos_sin_cache[compressed_pos, pair_idx] |
| 303 | sin_v = cos_sin_cache[ |
| 304 | compressed_pos, pair_idx + Int32(rope_dim // 2) |
| 305 | ] |
| 306 | real = x[elem] * cos_v - x[elem + 1] * sin_v |
| 307 | imag = x[elem] * sin_v + x[elem + 1] * cos_v |
| 308 | packed = _fp32x2_to_bf16x2(real, imag) |
| 309 | out_base = ( |
| 310 | value_base |
| 311 | + Int64(nope_dim) |
| 312 | + ((elem_base - nope_dim + Int32(elem)) * 2).to(Int64) |
| 313 | ) |
| 314 | k_cache_u32.iterator[out_base // Int64(4)] = packed |
| 315 | else: |
| 316 | q = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 317 | local_absmax = Float32(0.0) |
| 318 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 319 | elem = const_expr(pair * 2) |
| 320 | q_packed = _fp32x2_to_bf16x2(x[elem], x[elem + 1]) |
| 321 | q0, q1 = _bf16x2_to_fp32(q_packed) |
| 322 | q[elem] = q0 |
| 323 | q[elem + 1] = q1 |
| 324 | local_absmax = cute.arch.fmax( |
| 325 | local_absmax, |
| 326 | cute.arch.fmax(cute.math.absf(q0), cute.math.absf(q1)), |
| 327 | ) |
| 328 | absmax = local_absmax |
| 329 | group_mask_and_clamp = const_expr( |
| 330 | (cute.arch.WARP_SIZE - lanes_per_group) << 8 |
| 331 | | (cute.arch.WARP_SIZE - 1) |
| 332 | ) |
| 333 | for step in cutlass.range_constexpr(scale_reduce_steps): |
| 334 | offset = const_expr(scale_reduce_offset >> step) |
| 335 | absmax = cute.arch.fmax( |
| 336 | absmax, |
| 337 | cute.arch.shuffle_sync_bfly( |
| 338 | absmax, |
| 339 | offset=offset, |
| 340 | mask_and_clamp=group_mask_and_clamp, |
| 341 | ), |
| 342 | ) |
| 343 | scale_raw = cute.arch.fmax( |
| 344 | Float32(min_scale), |
| 345 | absmax / Float32(fp8_max), |
| 346 | ) |
| 347 | bits = _recast_val(scale_raw, Uint32) |
| 348 | ue8m0 = ((bits + Uint32(0x7FFFFF)) >> Uint32(23)) & Uint32(0xFF) |
| 349 | inv_scale = _recast_val( |
| 350 | (Uint32(254) - ue8m0) << Uint32(23), Float32 |
| 351 | ) |
| 352 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 353 | elem = const_expr(pair * 2) |
| 354 | y0 = cutlass.min( |
| 355 | cute.arch.fmax(q[elem] * inv_scale, Float32(-fp8_max)), |
| 356 | Float32(fp8_max), |
| 357 | ) |
| 358 | y1 = cutlass.min( |
| 359 | cute.arch.fmax(q[elem + 1] * inv_scale, Float32(-fp8_max)), |
| 360 | Float32(fp8_max), |
| 361 | ) |
| 362 | packed_fp8 = _fp32x2_to_fp8e4m3x2(y0, y1) |
| 363 | out_base = value_base + (elem_base + Int32(elem)).to(Int64) |
| 364 | k_cache_u16.iterator[out_base // Int64(2)] = packed_fp8 |
| 365 | if group_lane == 0: |
| 366 | k_cache.iterator[scale_base + group_idx.to(Int64)] = ue8m0.to( |
| 367 | Uint8 |
| 368 | ) |
| 369 | if group_idx == 0: |
| 370 | k_cache.iterator[scale_base + Int64(nope_blocks)] = Uint8(0) |
| 371 | |
| 372 | @cute.jit |
| 373 | def host_entrypoint( |
| 374 | state_cache: cute.Tensor, |
| 375 | token_to_req_indices: cute.Tensor, |
| 376 | positions: cute.Tensor, |
| 377 | slot_mapping: cute.Tensor, |
| 378 | block_table: cute.Tensor, |
| 379 | block_size: Int64, |
| 380 | rms_norm_weight: cute.Tensor, |
| 381 | rms_norm_eps: Float32, |
| 382 | cos_sin_cache: cute.Tensor, |
| 383 | k_cache: cute.Tensor, |
| 384 | kv_slot_mapping: cute.Tensor, |
| 385 | kv_cache_block_size: Int64, |
| 386 | stream: CUstream, |
| 387 | ): |
| 388 | grid = (slot_mapping.shape[0], 1, 1) |
| 389 | device_kernel( |
| 390 | state_cache, |
| 391 | token_to_req_indices, |
| 392 | positions, |
| 393 | slot_mapping, |
| 394 | block_table, |
| 395 | block_size, |
| 396 | rms_norm_weight, |
| 397 | rms_norm_eps, |
| 398 | cos_sin_cache, |
| 399 | k_cache, |
| 400 | kv_slot_mapping, |
| 401 | kv_cache_block_size, |
| 402 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 403 | |
| 404 | return host_entrypoint |
| 405 | |
| 406 | def dispatch( # type: ignore[override] |
| 407 | self, |
| 408 | *, |
| 409 | compress_ratio: int, |
| 410 | norm_weight_dtype: type[cutlass.Numeric], |
| 411 | head_size: int, |
| 412 | rope_head_dim: int, |
| 413 | ) -> CompileKey: |
| 414 | overlap = compress_ratio == 4 |
| 415 | state_width = head_size * (2 if overlap else 1) |
| 416 | return self.CompileKey( |
| 417 | head_size=head_size, |
| 418 | state_width=state_width, |
| 419 | rope_head_dim=rope_head_dim, |
| 420 | fp8_max=448.0, |
| 421 | quant_block=64, |
| 422 | token_stride=576, |
| 423 | scale_dim=8, |
| 424 | compress_ratio=compress_ratio, |
| 425 | overlap=overlap, |
| 426 | norm_weight_dtype=norm_weight_dtype, |
| 427 | ) |
| 428 | |
| 429 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 430 | hf_config = vllm_config.model_config.hf_config |
| 431 | cache_config = vllm_config.cache_config |
| 432 | return self._trace_dispatch(self.dispatch)( |
| 433 | compress_ratio=4, |
| 434 | norm_weight_dtype=torch_to_cute_dtype(vllm_config.model_config.dtype), |
| 435 | head_size=hf_config.head_dim, |
| 436 | rope_head_dim=hf_config.qk_rope_head_dim, |
| 437 | enabled=( |
| 438 | cache_config.block_size > 0 |
| 439 | and 4 in hf_config.compress_ratios |
| 440 | and cache_config.cache_dtype == "fp8_ds_mla" |
| 441 | ), |
| 442 | _when=lambda *, enabled: enabled, |
| 443 | ) |
| 444 | |
| 445 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 446 | head_size = compile_key.head_size |
| 447 | state_width = compile_key.state_width |
| 448 | rope_head_dim = compile_key.rope_head_dim |
| 449 | quant_block = compile_key.quant_block |
| 450 | token_stride = compile_key.token_stride |
| 451 | scale_dim = compile_key.scale_dim |
| 452 | compress_ratio = compile_key.compress_ratio |
| 453 | overlap = compile_key.overlap |
| 454 | norm_weight_dtype = compile_key.norm_weight_dtype |
| 455 | if compress_ratio != 4 or not overlap: |
| 456 | raise ValueError("CuTe DSL C4 fused sparse-attn requires C4 overlap.") |
| 457 | if head_size != 512: |
| 458 | raise ValueError( |
| 459 | "CuTe DSL C4 fused sparse-attn currently requires head_size=512.", |
| 460 | ) |
| 461 | if state_width != 2 * head_size: |
| 462 | raise ValueError( |
| 463 | "CuTe DSL C4 fused sparse-attn requires state_width=2*head_size.", |
| 464 | ) |
| 465 | if quant_block != 64: |
| 466 | raise ValueError( |
| 467 | "CuTe DSL C4 fused sparse-attn currently requires quant_block=64.", |
| 468 | ) |
| 469 | if rope_head_dim != 64: |
| 470 | raise ValueError( |
| 471 | "CuTe DSL C4 fused sparse-attn currently requires rope_head_dim=64.", |
| 472 | ) |
| 473 | if token_stride < head_size + rope_head_dim: |
| 474 | raise ValueError("token_stride is too small for the packed FP8/BF16 row.") |
| 475 | expected_scale_dim = (head_size - rope_head_dim) // quant_block + 1 |
| 476 | if scale_dim < expected_scale_dim: |
| 477 | raise ValueError("scale_dim is too small for the UE8M0 scale row.") |
| 478 | num_positions = cute.sym_int() |
| 479 | num_slots = cute.sym_int() |
| 480 | num_req_indices = cute.sym_int() |
| 481 | num_kv_slots = cute.sym_int() |
| 482 | num_state_blocks = cute.sym_int() |
| 483 | num_kv_blocks = cute.sym_int() |
| 484 | state_cache_block_size = cute.sym_int() |
| 485 | block_table_width = cute.sym_int() |
| 486 | max_pos = cute.sym_int() |
| 487 | state_cache_width = state_width * 2 |
| 488 | state_cache = cute.runtime.make_fake_tensor( |
| 489 | Float32, |
| 490 | (num_state_blocks, state_cache_block_size, state_cache_width), |
| 491 | stride=( |
| 492 | cute.sym_int64(divisibility=16), |
| 493 | cute.sym_int64(divisibility=16), |
| 494 | 1, |
| 495 | ), |
| 496 | assumed_align=16, |
| 497 | ) |
| 498 | token_to_req_indices = make_fake_tensor( |
| 499 | Int32, |
| 500 | (num_req_indices,), |
| 501 | divisibility=4, |
| 502 | ) |
| 503 | positions = make_fake_tensor(Int64, (num_positions,), divisibility=8) |
| 504 | slot_mapping = make_fake_tensor(Int64, (num_slots,), divisibility=8) |
| 505 | block_table = make_fake_tensor( |
| 506 | Int32, |
| 507 | (cute.sym_int(), block_table_width), |
| 508 | divisibility=1, |
| 509 | ) |
| 510 | rms_norm_weight = make_fake_tensor( |
| 511 | norm_weight_dtype, |
| 512 | (head_size,), |
| 513 | divisibility=4, |
| 514 | ) |
| 515 | cos_sin_cache = cute.runtime.make_fake_tensor( |
| 516 | Float32, |
| 517 | (max_pos, rope_head_dim), |
| 518 | stride=(cute.sym_int64(divisibility=4), 1), |
| 519 | assumed_align=4, |
| 520 | ) |
| 521 | k_cache = cute.runtime.make_fake_tensor( |
| 522 | Uint8, |
| 523 | (num_kv_blocks, cute.sym_int(), cute.sym_int()), |
| 524 | stride=(cute.sym_int64(divisibility=16), cute.sym_int64(divisibility=8), 1), |
| 525 | assumed_align=16, |
| 526 | ) |
| 527 | kv_slot_mapping = make_fake_tensor(Int64, (num_kv_slots,), divisibility=8) |
| 528 | |
| 529 | return ( |
| 530 | state_cache, |
| 531 | token_to_req_indices, |
| 532 | positions, |
| 533 | slot_mapping, |
| 534 | block_table, |
| 535 | Int64(0), |
| 536 | rms_norm_weight, |
| 537 | Float32(0.0), |
| 538 | cos_sin_cache, |
| 539 | k_cache, |
| 540 | kv_slot_mapping, |
| 541 | Int64(0), |
| 542 | ) |
| 543 | |
| 544 | @kernel_launcher |
| 545 | def __call__( |
| 546 | self, |
| 547 | *, |
| 548 | state_cache: torch.Tensor, |
| 549 | token_to_req_indices: torch.Tensor, |
| 550 | positions: torch.Tensor, |
| 551 | slot_mapping: torch.Tensor, |
| 552 | block_table: torch.Tensor, |
| 553 | block_size: int, |
| 554 | rms_norm_weight: torch.Tensor, |
| 555 | rms_norm_eps: float, |
| 556 | cos_sin_cache: torch.Tensor, |
| 557 | kv_cache: torch.Tensor, |
| 558 | kv_slot_mapping: torch.Tensor, |
| 559 | compress_ratio: int, |
| 560 | head_dim: int, |
| 561 | rope_head_dim: int, |
| 562 | ) -> CuTeDSLLaunchSpec[SparseAttnCompressNormRopeStoreC4Kernel.CompileKey]: |
| 563 | compile_key = self.dispatch( |
| 564 | compress_ratio=compress_ratio, |
| 565 | norm_weight_dtype=torch_to_cute_dtype(rms_norm_weight.dtype), |
| 566 | head_size=head_dim, |
| 567 | rope_head_dim=rope_head_dim, |
| 568 | ) |
| 569 | launch_args = ( |
| 570 | state_cache, |
| 571 | token_to_req_indices, |
| 572 | positions, |
| 573 | slot_mapping, |
| 574 | block_table, |
| 575 | block_size, |
| 576 | rms_norm_weight, |
| 577 | rms_norm_eps, |
| 578 | cos_sin_cache, |
| 579 | kv_cache, |
| 580 | kv_slot_mapping, |
| 581 | kv_cache.shape[1], |
| 582 | ) |
| 583 | return compile_key, launch_args |
| 584 | |
| 585 | |
| 586 | class SparseAttnCompressNormRopeStoreFullC4Kernel( |
| 587 | VllmCuTeDSLJitKernel["SparseAttnCompressNormRopeStoreFullC4Kernel.CompileKey"] |
| 588 | ): |
| 589 | rcp_ln2 = 1.4426950408889634 |
| 590 | |
| 591 | @dataclass(frozen=True) |
| 592 | class CompileKey: |
| 593 | head_size: int |
| 594 | state_width: int |
| 595 | rope_head_dim: int |
| 596 | fp8_max: float |
| 597 | quant_block: int |
| 598 | compress_ratio: int |
| 599 | overlap: bool |
| 600 | store_full_fp8: bool |
| 601 | norm_weight_dtype: type[cutlass.Numeric] |
| 602 | |
| 603 | @staticmethod |
| 604 | def kernel(compile_key: CompileKey) -> Any: |
| 605 | rcp_ln2 = SparseAttnCompressNormRopeStoreFullC4Kernel.rcp_ln2 |
| 606 | head_dim = compile_key.head_size |
| 607 | state_width = compile_key.state_width |
| 608 | rope_dim = compile_key.rope_head_dim |
| 609 | nope_dim = head_dim - rope_dim |
| 610 | fp8_max = compile_key.fp8_max |
| 611 | quant_block = compile_key.quant_block |
| 612 | elems_per_lane = 8 |
| 613 | copy_elems = 4 |
| 614 | copy_chunks = elems_per_lane // copy_elems |
| 615 | lanes_per_group = quant_block // elems_per_lane |
| 616 | groups_per_warp = 32 // lanes_per_group |
| 617 | num_warps = (head_dim // quant_block) // groups_per_warp |
| 618 | nope_blocks = nope_dim // quant_block |
| 619 | tb_size = num_warps * 32 |
| 620 | compress_ratio = compile_key.compress_ratio |
| 621 | overlap = compile_key.overlap |
| 622 | window = (1 + int(overlap)) * compress_ratio |
| 623 | store_full_fp8 = compile_key.store_full_fp8 |
| 624 | |
| 625 | @cute.kernel |
| 626 | def device_kernel( |
| 627 | state_cache: cute.Tensor, |
| 628 | token_to_req_indices: cute.Tensor, |
| 629 | positions: cute.Tensor, |
| 630 | slot_mapping: cute.Tensor, |
| 631 | block_table: cute.Tensor, |
| 632 | block_size: Int64, |
| 633 | rms_norm_weight: cute.Tensor, |
| 634 | rms_norm_eps: Float32, |
| 635 | cos_sin_cache: cute.Tensor, |
| 636 | k_cache: cute.Tensor, |
| 637 | kv_slot_mapping: cute.Tensor, |
| 638 | kv_cache_block_size: Int64, |
| 639 | fp8_scale: cute.Tensor, |
| 640 | ): |
| 641 | token_idx, _, _ = cute.arch.block_idx() |
| 642 | tid, _, _ = cute.arch.thread_idx() |
| 643 | warp_id = cute.arch.make_warp_uniform(tid // 32) |
| 644 | lane_id = tid % 32 |
| 645 | group_lane = lane_id % lanes_per_group |
| 646 | group_idx = warp_id * groups_per_warp + lane_id // lanes_per_group |
| 647 | elem_base = group_idx * quant_block + group_lane * elems_per_lane |
| 648 | |
| 649 | slot_id = slot_mapping[token_idx] |
| 650 | has_position = token_idx < positions.shape[0] |
| 651 | position = Int64(0) |
| 652 | if has_position: |
| 653 | position = positions[token_idx] |
| 654 | boundary = has_position and ( |
| 655 | (position + Int64(1)) % Int64(compress_ratio) == Int64(0) |
| 656 | ) |
| 657 | has_req_idx = token_idx < token_to_req_indices.shape[0] |
| 658 | has_kv_slot_idx = token_idx < kv_slot_mapping.shape[0] |
| 659 | kv_slot_idx = Int64(-1) |
| 660 | if has_kv_slot_idx: |
| 661 | kv_slot_idx = kv_slot_mapping[token_idx] |
| 662 | active = ( |
| 663 | slot_id >= Int64(0) |
| 664 | and has_req_idx |
| 665 | and boundary |
| 666 | and kv_slot_idx >= Int64(0) |
| 667 | ) |
| 668 | |
| 669 | if active: |
| 670 | req_idx = token_to_req_indices[token_idx] |
| 671 | start = position - Int64(window - 1) |
| 672 | |
| 673 | smem = cutlass.utils.SmemAllocator() |
| 674 | s_block_numbers = smem.allocate_tensor( |
| 675 | Int32, cute.make_layout((window,)), byte_alignment=4 |
| 676 | ) |
| 677 | partial_sums = smem.allocate_tensor( |
| 678 | Float32, cute.make_layout((num_warps,)), byte_alignment=4 |
| 679 | ) |
| 680 | rrms_shared = smem.allocate_tensor( |
| 681 | Float32, cute.make_layout((1,)), byte_alignment=4 |
| 682 | ) |
| 683 | |
| 684 | for row in cutlass.range_constexpr(window): |
| 685 | pos = start + Int64(row) |
| 686 | if tid == row: |
| 687 | block_number_i32 = Int32(0) |
| 688 | if pos >= Int64(0): |
| 689 | block_index = pos // block_size |
| 690 | block_number_i32 = block_table[req_idx, block_index] |
| 691 | s_block_numbers[row] = block_number_i32 |
| 692 | cute.arch.sync_threads() |
| 693 | |
| 694 | local_max = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 695 | local_sum = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 696 | local_product = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 697 | |
| 698 | for e in cutlass.range_constexpr(elems_per_lane): |
| 699 | local_max[e] = -Float32.inf |
| 700 | local_sum[e] = Float32(0.0) |
| 701 | local_product[e] = Float32(0.0) |
| 702 | |
| 703 | cp_f32x4 = cute.make_copy_atom( |
| 704 | cute.nvgpu.CopyUniversalOp(), Float32, num_bits_per_copy=128 |
| 705 | ) |
| 706 | copy_layout = cute.make_layout( |
| 707 | (copy_chunks, copy_elems), |
| 708 | stride=(copy_elems, 1), |
| 709 | ) |
| 710 | kv_vals = cute.make_rmem_tensor(copy_layout, Float32) |
| 711 | score_vals = cute.make_rmem_tensor(copy_layout, Float32) |
| 712 | |
| 713 | for row in cutlass.range_constexpr(window): |
| 714 | pos = start + Int64(row) |
| 715 | if pos >= Int64(0): |
| 716 | block_index = pos // block_size |
| 717 | block_offset = pos - block_index * block_size |
| 718 | block_number = s_block_numbers[row].to(Int64) |
| 719 | head_offset = Int64((row // compress_ratio) * head_dim) |
| 720 | row_tensor = state_cache[block_number, block_offset, None] |
| 721 | for chunk in cutlass.range_constexpr(copy_chunks): |
| 722 | copy_elem = const_expr(chunk * copy_elems) |
| 723 | col_tile = ( |
| 724 | head_offset + (elem_base + Int32(copy_elem)).to(Int64) |
| 725 | ) // Int64(copy_elems) |
| 726 | kv_src = cute.local_tile( |
| 727 | row_tensor, |
| 728 | tiler=(copy_elems,), |
| 729 | coord=(col_tile,), |
| 730 | ) |
| 731 | score_src = cute.local_tile( |
| 732 | row_tensor, |
| 733 | tiler=(copy_elems,), |
| 734 | coord=(col_tile + Int64(state_width // copy_elems),), |
| 735 | ) |
| 736 | cute.copy(cp_f32x4, kv_src, kv_vals[chunk, None]) |
| 737 | cute.copy(cp_f32x4, score_src, score_vals[chunk, None]) |
| 738 | |
| 739 | for e in cutlass.range_constexpr(elems_per_lane): |
| 740 | chunk = const_expr(e // copy_elems) |
| 741 | copy_elem = const_expr(e % copy_elems) |
| 742 | score = score_vals[chunk, copy_elem] |
| 743 | kv = kv_vals[chunk, copy_elem] |
| 744 | new_max = cute.arch.fmax(local_max[e], score) |
| 745 | old_scale = cute.math.exp2( |
| 746 | (local_max[e] - new_max) * Float32(rcp_ln2), |
| 747 | fastmath=True, |
| 748 | ) |
| 749 | new_scale = cute.math.exp2( |
| 750 | (score - new_max) * Float32(rcp_ln2), |
| 751 | fastmath=True, |
| 752 | ) |
| 753 | local_sum[e] = local_sum[e] * old_scale + new_scale |
| 754 | local_product[e] = ( |
| 755 | local_product[e] * old_scale + kv * new_scale |
| 756 | ) |
| 757 | local_max[e] = new_max |
| 758 | |
| 759 | x = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 760 | local_sumsq = Float32(0.0) |
| 761 | for e in cutlass.range_constexpr(elems_per_lane): |
| 762 | x[e] = local_product[e] / local_sum[e] |
| 763 | local_sumsq += x[e] * x[e] |
| 764 | |
| 765 | warp_sum = local_sumsq |
| 766 | for step in cutlass.range_constexpr(5): |
| 767 | offset = const_expr(16 >> step) |
| 768 | warp_sum += cute.arch.shuffle_sync_bfly(warp_sum, offset) |
| 769 | |
| 770 | if lane_id == 0: |
| 771 | partial_sums[warp_id] = warp_sum |
| 772 | cute.arch.sync_threads() |
| 773 | if tid == 0: |
| 774 | total = Float32(0.0) |
| 775 | for i in cutlass.range_constexpr(num_warps): |
| 776 | total += partial_sums[i] |
| 777 | rrms_shared[0] = cute.math.rsqrt( |
| 778 | total / Float32(head_dim) + rms_norm_eps, fastmath=True |
| 779 | ) |
| 780 | cute.arch.sync_threads() |
| 781 | |
| 782 | rrms = rrms_shared[0] |
| 783 | for e in cutlass.range_constexpr(elems_per_lane): |
| 784 | elem = elem_base + e |
| 785 | x[e] = x[e] * rrms * rms_norm_weight[elem].to(Float32) |
| 786 | |
| 787 | page = kv_slot_idx // kv_cache_block_size |
| 788 | kv_offset = kv_slot_idx - page * kv_cache_block_size |
| 789 | value_base = page * k_cache.stride[0] + kv_offset * k_cache.stride[1] |
| 790 | |
| 791 | if const_expr(store_full_fp8): |
| 792 | k_cache_u16 = cute.recast_tensor(k_cache, Uint16) |
| 793 | inv_fp8 = Float32(1.0) / fp8_scale[0] |
| 794 | if group_idx == nope_blocks: |
| 795 | compressed_pos = (position // Int64(compress_ratio)) * Int64( |
| 796 | compress_ratio |
| 797 | ) |
| 798 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 799 | elem = const_expr(pair * 2) |
| 800 | pair_idx = (elem_base - nope_dim) // 2 + Int32(pair) |
| 801 | cos_v = cos_sin_cache[compressed_pos, pair_idx] |
| 802 | sin_v = cos_sin_cache[ |
| 803 | compressed_pos, pair_idx + Int32(rope_dim // 2) |
| 804 | ] |
| 805 | real = x[elem] * cos_v - x[elem + 1] * sin_v |
| 806 | imag = x[elem] * sin_v + x[elem + 1] * cos_v |
| 807 | packed_bf16 = _fp32x2_to_bf16x2(real, imag) |
| 808 | b0, b1 = _bf16x2_to_fp32(packed_bf16) |
| 809 | y0 = cutlass.min( |
| 810 | cutlass.max(b0 * inv_fp8, Float32(-fp8_max)), |
| 811 | Float32(fp8_max), |
| 812 | ) |
| 813 | y1 = cutlass.min( |
| 814 | cutlass.max(b1 * inv_fp8, Float32(-fp8_max)), |
| 815 | Float32(fp8_max), |
| 816 | ) |
| 817 | packed_fp8 = _fp32x2_to_fp8e4m3x2(y0, y1) |
| 818 | out_base = value_base + (elem_base + Int32(elem)).to(Int64) |
| 819 | k_cache_u16.iterator[out_base // Int64(2)] = packed_fp8 |
| 820 | else: |
| 821 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 822 | elem = const_expr(pair * 2) |
| 823 | packed_bf16 = _fp32x2_to_bf16x2(x[elem], x[elem + 1]) |
| 824 | b0, b1 = _bf16x2_to_fp32(packed_bf16) |
| 825 | y0 = cutlass.min( |
| 826 | cutlass.max(b0 * inv_fp8, Float32(-fp8_max)), |
| 827 | Float32(fp8_max), |
| 828 | ) |
| 829 | y1 = cutlass.min( |
| 830 | cutlass.max(b1 * inv_fp8, Float32(-fp8_max)), |
| 831 | Float32(fp8_max), |
| 832 | ) |
| 833 | packed_fp8 = _fp32x2_to_fp8e4m3x2(y0, y1) |
| 834 | out_base = value_base + (elem_base + Int32(elem)).to(Int64) |
| 835 | k_cache_u16.iterator[out_base // Int64(2)] = packed_fp8 |
| 836 | else: |
| 837 | k_cache_u32 = cute.recast_tensor(k_cache, Uint32) |
| 838 | if group_idx == nope_blocks: |
| 839 | compressed_pos = (position // Int64(compress_ratio)) * Int64( |
| 840 | compress_ratio |
| 841 | ) |
| 842 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 843 | elem = const_expr(pair * 2) |
| 844 | pair_idx = (elem_base - nope_dim) // 2 + Int32(pair) |
| 845 | cos_v = cos_sin_cache[compressed_pos, pair_idx] |
| 846 | sin_v = cos_sin_cache[ |
| 847 | compressed_pos, pair_idx + Int32(rope_dim // 2) |
| 848 | ] |
| 849 | real = x[elem] * cos_v - x[elem + 1] * sin_v |
| 850 | imag = x[elem] * sin_v + x[elem + 1] * cos_v |
| 851 | packed_bf16 = _fp32x2_to_bf16x2(real, imag) |
| 852 | out_base = value_base + ((elem_base + Int32(elem)) * 2).to( |
| 853 | Int64 |
| 854 | ) |
| 855 | k_cache_u32.iterator[out_base // Int64(4)] = packed_bf16 |
| 856 | else: |
| 857 | for pair in cutlass.range_constexpr(elems_per_lane // 2): |
| 858 | elem = const_expr(pair * 2) |
| 859 | packed_bf16 = _fp32x2_to_bf16x2(x[elem], x[elem + 1]) |
| 860 | out_base = value_base + ((elem_base + Int32(elem)) * 2).to( |
| 861 | Int64 |
| 862 | ) |
| 863 | k_cache_u32.iterator[out_base // Int64(4)] = packed_bf16 |
| 864 | |
| 865 | @cute.jit |
| 866 | def host_entrypoint( |
| 867 | state_cache: cute.Tensor, |
| 868 | token_to_req_indices: cute.Tensor, |
| 869 | positions: cute.Tensor, |
| 870 | slot_mapping: cute.Tensor, |
| 871 | block_table: cute.Tensor, |
| 872 | block_size: Int64, |
| 873 | rms_norm_weight: cute.Tensor, |
| 874 | rms_norm_eps: Float32, |
| 875 | cos_sin_cache: cute.Tensor, |
| 876 | k_cache: cute.Tensor, |
| 877 | kv_slot_mapping: cute.Tensor, |
| 878 | kv_cache_block_size: Int64, |
| 879 | fp8_scale: cute.Tensor, |
| 880 | stream: CUstream, |
| 881 | ): |
| 882 | grid = (slot_mapping.shape[0], 1, 1) |
| 883 | device_kernel( |
| 884 | state_cache, |
| 885 | token_to_req_indices, |
| 886 | positions, |
| 887 | slot_mapping, |
| 888 | block_table, |
| 889 | block_size, |
| 890 | rms_norm_weight, |
| 891 | rms_norm_eps, |
| 892 | cos_sin_cache, |
| 893 | k_cache, |
| 894 | kv_slot_mapping, |
| 895 | kv_cache_block_size, |
| 896 | fp8_scale, |
| 897 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 898 | |
| 899 | return host_entrypoint |
| 900 | |
| 901 | def dispatch( # type: ignore[override] |
| 902 | self, |
| 903 | *, |
| 904 | compress_ratio: int, |
| 905 | store_full_fp8: bool, |
| 906 | norm_weight_dtype: type[cutlass.Numeric], |
| 907 | head_size: int, |
| 908 | rope_head_dim: int, |
| 909 | ) -> CompileKey: |
| 910 | overlap = compress_ratio == 4 |
| 911 | return self.CompileKey( |
| 912 | head_size=head_size, |
| 913 | state_width=head_size * (2 if overlap else 1), |
| 914 | rope_head_dim=rope_head_dim, |
| 915 | fp8_max=448.0, |
| 916 | quant_block=64, |
| 917 | compress_ratio=compress_ratio, |
| 918 | overlap=overlap, |
| 919 | store_full_fp8=store_full_fp8, |
| 920 | norm_weight_dtype=norm_weight_dtype, |
| 921 | ) |
| 922 | |
| 923 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 924 | hf_config = vllm_config.model_config.hf_config |
| 925 | cache_config = vllm_config.cache_config |
| 926 | return self._trace_dispatch(self.dispatch)( |
| 927 | compress_ratio=4, |
| 928 | store_full_fp8=( |
| 929 | cache_config.cache_dtype.startswith("fp8") |
| 930 | and cache_config.cache_dtype != "fp8_ds_mla" |
| 931 | ), |
| 932 | norm_weight_dtype=torch_to_cute_dtype(vllm_config.model_config.dtype), |
| 933 | head_size=hf_config.head_dim, |
| 934 | rope_head_dim=hf_config.qk_rope_head_dim, |
| 935 | enabled=( |
| 936 | cache_config.block_size > 0 |
| 937 | and 4 in hf_config.compress_ratios |
| 938 | and cache_config.cache_dtype != "fp8_ds_mla" |
| 939 | ), |
| 940 | _when=lambda *, enabled: enabled, |
| 941 | ) |
| 942 | |
| 943 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 944 | head_size = compile_key.head_size |
| 945 | state_width = compile_key.state_width |
| 946 | rope_head_dim = compile_key.rope_head_dim |
| 947 | quant_block = compile_key.quant_block |
| 948 | compress_ratio = compile_key.compress_ratio |
| 949 | overlap = compile_key.overlap |
| 950 | norm_weight_dtype = compile_key.norm_weight_dtype |
| 951 | if compress_ratio != 4 or not overlap: |
| 952 | raise ValueError("CuTe DSL C4 fused sparse-attn requires C4 overlap.") |
| 953 | if head_size != 512: |
| 954 | raise ValueError( |
| 955 | "CuTe DSL C4 fused sparse-attn currently requires head_size=512.", |
| 956 | ) |
| 957 | if state_width != 2 * head_size: |
| 958 | raise ValueError( |
| 959 | "CuTe DSL C4 fused sparse-attn requires state_width=2*head_size.", |
| 960 | ) |
| 961 | if quant_block != 64: |
| 962 | raise ValueError( |
| 963 | "CuTe DSL C4 fused sparse-attn currently requires quant_block=64.", |
| 964 | ) |
| 965 | if rope_head_dim != 64: |
| 966 | raise ValueError( |
| 967 | "CuTe DSL C4 fused sparse-attn currently requires rope_head_dim=64.", |
| 968 | ) |
| 969 | num_positions = cute.sym_int() |
| 970 | num_slots = cute.sym_int() |
| 971 | num_req_indices = cute.sym_int() |
| 972 | num_kv_slots = cute.sym_int() |
| 973 | num_state_blocks = cute.sym_int() |
| 974 | num_kv_blocks = cute.sym_int() |
| 975 | state_cache_block_size = cute.sym_int() |
| 976 | block_table_width = cute.sym_int() |
| 977 | max_pos = cute.sym_int() |
| 978 | state_cache_width = state_width * 2 |
| 979 | state_cache = cute.runtime.make_fake_tensor( |
| 980 | Float32, |
| 981 | (num_state_blocks, state_cache_block_size, state_cache_width), |
| 982 | stride=( |
| 983 | cute.sym_int64(divisibility=16), |
| 984 | cute.sym_int64(divisibility=16), |
| 985 | 1, |
| 986 | ), |
| 987 | assumed_align=16, |
| 988 | ) |
| 989 | token_to_req_indices = make_fake_tensor( |
| 990 | Int32, |
| 991 | (num_req_indices,), |
| 992 | divisibility=4, |
| 993 | ) |
| 994 | positions = make_fake_tensor(Int64, (num_positions,), divisibility=8) |
| 995 | slot_mapping = make_fake_tensor(Int64, (num_slots,), divisibility=8) |
| 996 | block_table = make_fake_tensor( |
| 997 | Int32, |
| 998 | (cute.sym_int(), block_table_width), |
| 999 | divisibility=1, |
| 1000 | ) |
| 1001 | rms_norm_weight = make_fake_tensor( |
| 1002 | norm_weight_dtype, |
| 1003 | (head_size,), |
| 1004 | divisibility=4, |
| 1005 | ) |
| 1006 | cos_sin_cache = cute.runtime.make_fake_tensor( |
| 1007 | Float32, |
| 1008 | (max_pos, rope_head_dim), |
| 1009 | stride=(cute.sym_int64(divisibility=4), 1), |
| 1010 | assumed_align=4, |
| 1011 | ) |
| 1012 | k_cache = cute.runtime.make_fake_tensor( |
| 1013 | Uint8, |
| 1014 | (num_kv_blocks, cute.sym_int(), cute.sym_int()), |
| 1015 | stride=(cute.sym_int64(divisibility=16), cute.sym_int64(divisibility=8), 1), |
| 1016 | assumed_align=16, |
| 1017 | ) |
| 1018 | kv_slot_mapping = make_fake_tensor(Int64, (num_kv_slots,), divisibility=8) |
| 1019 | fp8_scale = make_fake_tensor(Float32, (1,), divisibility=1) |
| 1020 | return ( |
| 1021 | state_cache, |
| 1022 | token_to_req_indices, |
| 1023 | positions, |
| 1024 | slot_mapping, |
| 1025 | block_table, |
| 1026 | Int64(0), |
| 1027 | rms_norm_weight, |
| 1028 | Float32(0.0), |
| 1029 | cos_sin_cache, |
| 1030 | k_cache, |
| 1031 | kv_slot_mapping, |
| 1032 | Int64(0), |
| 1033 | fp8_scale, |
| 1034 | ) |
| 1035 | |
| 1036 | @kernel_launcher |
| 1037 | def __call__( |
| 1038 | self, |
| 1039 | *, |
| 1040 | state_cache: torch.Tensor, |
| 1041 | token_to_req_indices: torch.Tensor, |
| 1042 | positions: torch.Tensor, |
| 1043 | slot_mapping: torch.Tensor, |
| 1044 | block_table: torch.Tensor, |
| 1045 | block_size: int, |
| 1046 | rms_norm_weight: torch.Tensor, |
| 1047 | rms_norm_eps: float, |
| 1048 | cos_sin_cache: torch.Tensor, |
| 1049 | kv_cache: torch.Tensor, |
| 1050 | kv_slot_mapping: torch.Tensor, |
| 1051 | compress_ratio: int, |
| 1052 | store_full_fp8: bool, |
| 1053 | fp8_scale: torch.Tensor | None, |
| 1054 | head_dim: int, |
| 1055 | rope_head_dim: int, |
| 1056 | ) -> CuTeDSLLaunchSpec[SparseAttnCompressNormRopeStoreFullC4Kernel.CompileKey]: |
| 1057 | compile_key = self.dispatch( |
| 1058 | compress_ratio=compress_ratio, |
| 1059 | store_full_fp8=store_full_fp8, |
| 1060 | norm_weight_dtype=torch_to_cute_dtype(rms_norm_weight.dtype), |
| 1061 | head_size=head_dim, |
| 1062 | rope_head_dim=rope_head_dim, |
| 1063 | ) |
| 1064 | if fp8_scale is None: |
| 1065 | fp8_scale = torch.ones(1, dtype=torch.float32, device=kv_cache.device) |
| 1066 | launch_args = ( |
| 1067 | state_cache, |
| 1068 | token_to_req_indices, |
| 1069 | positions, |
| 1070 | slot_mapping, |
| 1071 | block_table, |
| 1072 | block_size, |
| 1073 | rms_norm_weight, |
| 1074 | rms_norm_eps, |
| 1075 | cos_sin_cache, |
| 1076 | kv_cache.view(torch.uint8), |
| 1077 | kv_slot_mapping, |
| 1078 | kv_cache.shape[1], |
| 1079 | fp8_scale, |
| 1080 | ) |
| 1081 | return compile_key, launch_args |
| 1082 | |
| 1083 | |
| 1084 | class SparseAttnCompressC128Block8Kernel( |
| 1085 | VllmCuTeDSLJitKernel["SparseAttnCompressC128Block8Kernel.CompileKey"] |
| 1086 | ): |
| 1087 | head_tile = 64 |
| 1088 | rows_per_warp = 16 |
| 1089 | elems_per_lane = 2 |
| 1090 | lanes_per_row = head_tile // elems_per_lane |
| 1091 | num_warps = 8 |
| 1092 | stats_lane_stride = lanes_per_row + 1 |
| 1093 | final_reduce_steps = 3 |
| 1094 | final_reduce_initial_offset = 4 |
| 1095 | tb_size = num_warps * 32 |
| 1096 | compress_ratio = 128 |
| 1097 | state_block_size = 8 |
| 1098 | rcp_ln2 = 1.4426950408889634 |
| 1099 | |
| 1100 | @dataclass(frozen=True) |
| 1101 | class CompileKey: |
| 1102 | head_size: int |
| 1103 | state_width: int |
| 1104 | |
| 1105 | @staticmethod |
| 1106 | def kernel(compile_key: CompileKey) -> Any: |
| 1107 | head_tile = SparseAttnCompressC128Block8Kernel.head_tile |
| 1108 | rows_per_warp = SparseAttnCompressC128Block8Kernel.rows_per_warp |
| 1109 | elems_per_lane = SparseAttnCompressC128Block8Kernel.elems_per_lane |
| 1110 | lanes_per_row = SparseAttnCompressC128Block8Kernel.lanes_per_row |
| 1111 | num_warps = SparseAttnCompressC128Block8Kernel.num_warps |
| 1112 | stats_lane_stride = SparseAttnCompressC128Block8Kernel.stats_lane_stride |
| 1113 | final_reduce_steps = SparseAttnCompressC128Block8Kernel.final_reduce_steps |
| 1114 | final_reduce_initial_offset = ( |
| 1115 | SparseAttnCompressC128Block8Kernel.final_reduce_initial_offset |
| 1116 | ) |
| 1117 | tb_size = SparseAttnCompressC128Block8Kernel.tb_size |
| 1118 | compress_ratio = SparseAttnCompressC128Block8Kernel.compress_ratio |
| 1119 | state_block_size = SparseAttnCompressC128Block8Kernel.state_block_size |
| 1120 | rcp_ln2 = SparseAttnCompressC128Block8Kernel.rcp_ln2 |
| 1121 | head_dim = compile_key.head_size |
| 1122 | num_splits = head_dim // head_tile |
| 1123 | state_width = compile_key.state_width |
| 1124 | |
| 1125 | @cute.kernel |
| 1126 | def device_kernel( |
| 1127 | state_cache: cute.Tensor, |
| 1128 | token_to_req_indices: cute.Tensor, |
| 1129 | positions: cute.Tensor, |
| 1130 | slot_mapping: cute.Tensor, |
| 1131 | block_table: cute.Tensor, |
| 1132 | compressed_kv: cute.Tensor, |
| 1133 | ): |
| 1134 | block_id, _, _ = cute.arch.block_idx() |
| 1135 | tid, _, _ = cute.arch.thread_idx() |
| 1136 | warp_id = cute.arch.make_warp_uniform(tid // 32) |
| 1137 | lane_id = tid % 32 |
| 1138 | col_group = lane_id % lanes_per_row |
| 1139 | |
| 1140 | token_idx = block_id // num_splits |
| 1141 | split_idx = block_id - token_idx * num_splits |
| 1142 | col_base = split_idx * head_tile + col_group * elems_per_lane |
| 1143 | |
| 1144 | position = Int64(0) |
| 1145 | req_idx = Int32(0) |
| 1146 | slot_id = Int64(-1) |
| 1147 | has_position = token_idx < positions.shape[0] |
| 1148 | has_req_idx = token_idx < token_to_req_indices.shape[0] |
| 1149 | if lane_id == 0: |
| 1150 | slot_id = slot_mapping[token_idx] |
| 1151 | if lane_id == 0 and has_position: |
| 1152 | position = positions[token_idx] |
| 1153 | if lane_id == 0 and has_req_idx: |
| 1154 | req_idx = token_to_req_indices[token_idx] |
| 1155 | slot_id = cute.arch.shuffle_sync(slot_id, offset=0) |
| 1156 | position = cute.arch.shuffle_sync(position, offset=0) |
| 1157 | req_idx = cute.arch.shuffle_sync(req_idx, offset=0) |
| 1158 | boundary = has_position and ( |
| 1159 | (position + Int64(1)) % Int64(compress_ratio) == Int64(0) |
| 1160 | ) |
| 1161 | start = position - Int64(compress_ratio - 1) |
| 1162 | active = slot_id >= Int64(0) and has_req_idx and boundary |
| 1163 | |
| 1164 | if active: |
| 1165 | smem = cutlass.utils.SmemAllocator() |
| 1166 | s_max = smem.allocate_tensor( |
| 1167 | Float32, |
| 1168 | cute.make_layout( |
| 1169 | ( |
| 1170 | num_warps, |
| 1171 | lanes_per_row, |
| 1172 | elems_per_lane, |
| 1173 | ), |
| 1174 | stride=( |
| 1175 | stats_lane_stride * elems_per_lane, |
| 1176 | elems_per_lane, |
| 1177 | 1, |
| 1178 | ), |
| 1179 | ), |
| 1180 | byte_alignment=4, |
| 1181 | ) |
| 1182 | s_sum = smem.allocate_tensor( |
| 1183 | Float32, |
| 1184 | cute.make_layout( |
| 1185 | ( |
| 1186 | num_warps, |
| 1187 | lanes_per_row, |
| 1188 | elems_per_lane, |
| 1189 | ), |
| 1190 | stride=( |
| 1191 | stats_lane_stride * elems_per_lane, |
| 1192 | elems_per_lane, |
| 1193 | 1, |
| 1194 | ), |
| 1195 | ), |
| 1196 | byte_alignment=4, |
| 1197 | ) |
| 1198 | s_product = smem.allocate_tensor( |
| 1199 | Float32, |
| 1200 | cute.make_layout( |
| 1201 | ( |
| 1202 | num_warps, |
| 1203 | lanes_per_row, |
| 1204 | elems_per_lane, |
| 1205 | ), |
| 1206 | stride=( |
| 1207 | stats_lane_stride * elems_per_lane, |
| 1208 | elems_per_lane, |
| 1209 | 1, |
| 1210 | ), |
| 1211 | ), |
| 1212 | byte_alignment=4, |
| 1213 | ) |
| 1214 | |
| 1215 | row_layout = cute.make_layout( |
| 1216 | (rows_per_warp, elems_per_lane), |
| 1217 | stride=(elems_per_lane, 1), |
| 1218 | ) |
| 1219 | kv_vals = cute.make_rmem_tensor(row_layout, Float32) |
| 1220 | score_vals = cute.make_rmem_tensor(row_layout, Float32) |
| 1221 | local_max = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 1222 | local_sum = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 1223 | local_product = cute.make_rmem_tensor((elems_per_lane,), Float32) |
| 1224 | |
| 1225 | for e in cutlass.range_constexpr(elems_per_lane): |
| 1226 | local_max[e] = -Float32.inf |
| 1227 | local_sum[e] = Float32(0.0) |
| 1228 | local_product[e] = Float32(0.0) |
| 1229 | |
| 1230 | first_block_index = start // Int64(state_block_size) |
| 1231 | warp_block_index = first_block_index + (warp_id * 2).to(Int64) |
| 1232 | block0_i32 = Int32(0) |
| 1233 | block1_i32 = Int32(0) |
| 1234 | if lane_id == 0: |
| 1235 | block0_i32 = block_table[req_idx, warp_block_index] |
| 1236 | block1_i32 = block_table[req_idx, warp_block_index + Int64(1)] |
| 1237 | block0_i32 = cute.arch.shuffle_sync(block0_i32, offset=0) |
| 1238 | block1_i32 = cute.arch.shuffle_sync(block1_i32, offset=0) |
| 1239 | |
| 1240 | cp_f32x2 = cute.make_copy_atom( |
| 1241 | cute.nvgpu.CopyUniversalOp(), Float32, num_bits_per_copy=64 |
| 1242 | ) |
| 1243 | final_mask_and_clamp = const_expr( |
| 1244 | (cute.arch.WARP_SIZE - num_warps) << 8 | (cute.arch.WARP_SIZE - 1) |
| 1245 | ) |
| 1246 | col_tile = col_base.to(Int64) // Int64(elems_per_lane) |
| 1247 | score_col_tile = col_tile + Int64(state_width // elems_per_lane) |
| 1248 | |
| 1249 | for i in cutlass.range_constexpr(rows_per_warp): |
| 1250 | block_number_i32 = block0_i32 |
| 1251 | block_offset = Int64(i) |
| 1252 | if const_expr(i >= state_block_size): |
| 1253 | block_number_i32 = block1_i32 |
| 1254 | block_offset = Int64(i - state_block_size) |
| 1255 | row_tensor = state_cache[ |
| 1256 | block_number_i32.to(Int64), block_offset, None |
| 1257 | ] |
| 1258 | kv_src = cute.local_tile( |
| 1259 | row_tensor, |
| 1260 | tiler=(elems_per_lane,), |
| 1261 | coord=(col_tile,), |
| 1262 | ) |
| 1263 | score_src = cute.local_tile( |
| 1264 | row_tensor, |
| 1265 | tiler=(elems_per_lane,), |
| 1266 | coord=(score_col_tile,), |
| 1267 | ) |
| 1268 | cute.copy(cp_f32x2, kv_src, kv_vals[i, None]) |
| 1269 | cute.copy(cp_f32x2, score_src, score_vals[i, None]) |
| 1270 | |
| 1271 | for e in cutlass.range_constexpr(elems_per_lane): |
| 1272 | local_max[e] = cute.arch.fmax(local_max[e], score_vals[i, e]) |
| 1273 | |
| 1274 | for e in cutlass.range_constexpr(elems_per_lane): |
| 1275 | for i in cutlass.range_constexpr(rows_per_warp): |
| 1276 | exp_score = cute.math.exp2( |
| 1277 | (score_vals[i, e] - local_max[e]) * Float32(rcp_ln2), |
| 1278 | fastmath=True, |
| 1279 | ) |
| 1280 | local_sum[e] += exp_score |
| 1281 | local_product[e] += kv_vals[i, e] * exp_score |
| 1282 | |
| 1283 | for e in cutlass.range_constexpr(elems_per_lane): |
| 1284 | s_max[warp_id, col_group, e] = local_max[e] |
| 1285 | s_sum[warp_id, col_group, e] = local_sum[e] |
| 1286 | s_product[warp_id, col_group, e] = local_product[e] |
| 1287 | cute.arch.sync_threads() |
| 1288 | |
| 1289 | out_group = tid // num_warps |
| 1290 | final_lane = tid % num_warps |
| 1291 | final_groups_per_pass = const_expr(tb_size // num_warps) |
| 1292 | for pass_idx in cutlass.range_constexpr( |
| 1293 | head_tile // final_groups_per_pass |
| 1294 | ): |
| 1295 | out_idx = pass_idx * final_groups_per_pass + out_group |
| 1296 | out_lane = out_idx // elems_per_lane |
| 1297 | out_elem = out_idx % elems_per_lane |
| 1298 | |
| 1299 | local_warp_max = s_max[final_lane, out_lane, out_elem] |
| 1300 | global_max = local_warp_max |
| 1301 | for step in cutlass.range_constexpr(final_reduce_steps): |
| 1302 | offset = const_expr(final_reduce_initial_offset >> step) |
| 1303 | global_max = cute.arch.fmax( |
| 1304 | global_max, |
| 1305 | cute.arch.shuffle_sync_bfly( |
| 1306 | global_max, |
| 1307 | offset=offset, |
| 1308 | mask_and_clamp=final_mask_and_clamp, |
| 1309 | ), |
| 1310 | ) |
| 1311 | |
| 1312 | scale = cute.math.exp2( |
| 1313 | (local_warp_max - global_max) * Float32(rcp_ln2), |
| 1314 | fastmath=True, |
| 1315 | ) |
| 1316 | global_sum = s_sum[final_lane, out_lane, out_elem] * scale |
| 1317 | global_product = s_product[final_lane, out_lane, out_elem] * scale |
| 1318 | for step in cutlass.range_constexpr(final_reduce_steps): |
| 1319 | offset = const_expr(final_reduce_initial_offset >> step) |
| 1320 | global_sum += cute.arch.shuffle_sync_bfly( |
| 1321 | global_sum, |
| 1322 | offset=offset, |
| 1323 | mask_and_clamp=final_mask_and_clamp, |
| 1324 | ) |
| 1325 | global_product += cute.arch.shuffle_sync_bfly( |
| 1326 | global_product, |
| 1327 | offset=offset, |
| 1328 | mask_and_clamp=final_mask_and_clamp, |
| 1329 | ) |
| 1330 | |
| 1331 | if final_lane == 0: |
| 1332 | compressed_kv.iterator[ |
| 1333 | token_idx.to(Int64) * compressed_kv.stride[0] |
| 1334 | + (split_idx * head_tile + out_idx).to(Int64) |
| 1335 | ] = global_product / global_sum |
| 1336 | |
| 1337 | @cute.jit |
| 1338 | def host_entrypoint( |
| 1339 | state_cache: cute.Tensor, |
| 1340 | token_to_req_indices: cute.Tensor, |
| 1341 | positions: cute.Tensor, |
| 1342 | slot_mapping: cute.Tensor, |
| 1343 | block_table: cute.Tensor, |
| 1344 | compressed_kv: cute.Tensor, |
| 1345 | stream: CUstream, |
| 1346 | ): |
| 1347 | grid = (slot_mapping.shape[0] * num_splits, 1, 1) |
| 1348 | device_kernel( |
| 1349 | state_cache, |
| 1350 | token_to_req_indices, |
| 1351 | positions, |
| 1352 | slot_mapping, |
| 1353 | block_table, |
| 1354 | compressed_kv, |
| 1355 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 1356 | |
| 1357 | return host_entrypoint |
| 1358 | |
| 1359 | def dispatch( # type: ignore[override] |
| 1360 | self, *, head_size: int, state_width: int |
| 1361 | ) -> CompileKey: |
| 1362 | return self.CompileKey(head_size=head_size, state_width=state_width) |
| 1363 | |
| 1364 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 1365 | hf_config = vllm_config.model_config.hf_config |
| 1366 | return self._trace_dispatch(self.dispatch)( |
| 1367 | head_size=hf_config.head_dim, |
| 1368 | state_width=hf_config.head_dim, |
| 1369 | enabled=128 in hf_config.compress_ratios, |
| 1370 | _when=lambda *, enabled: enabled, |
| 1371 | ) |
| 1372 | |
| 1373 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 1374 | head_size = compile_key.head_size |
| 1375 | state_width = compile_key.state_width |
| 1376 | if head_size % SparseAttnCompressC128Block8Kernel.head_tile != 0: |
| 1377 | raise ValueError("head_size must be divisible by the 64-wide head tile.") |
| 1378 | num_positions = cute.sym_int() |
| 1379 | num_slots = cute.sym_int() |
| 1380 | num_req_indices = cute.sym_int() |
| 1381 | num_blocks = cute.sym_int() |
| 1382 | state_cache_block_size = cute.sym_int() |
| 1383 | block_table_width = cute.sym_int() |
| 1384 | state_cache_width = state_width * 2 |
| 1385 | state_cache = cute.runtime.make_fake_tensor( |
| 1386 | Float32, |
| 1387 | (num_blocks, state_cache_block_size, state_cache_width), |
| 1388 | stride=( |
| 1389 | cute.sym_int64(divisibility=16), |
| 1390 | cute.sym_int64(divisibility=16), |
| 1391 | 1, |
| 1392 | ), |
| 1393 | assumed_align=16, |
| 1394 | ) |
| 1395 | token_to_req_indices = make_fake_tensor( |
| 1396 | Int32, |
| 1397 | (num_req_indices,), |
| 1398 | divisibility=4, |
| 1399 | ) |
| 1400 | positions = make_fake_tensor(Int64, (num_positions,), divisibility=8) |
| 1401 | slot_mapping = make_fake_tensor(Int64, (num_slots,), divisibility=8) |
| 1402 | block_table = make_fake_tensor( |
| 1403 | Int32, |
| 1404 | (cute.sym_int(), block_table_width), |
| 1405 | divisibility=1, |
| 1406 | ) |
| 1407 | compressed_kv = cute.runtime.make_fake_tensor( |
| 1408 | Float32, |
| 1409 | (num_slots, head_size), |
| 1410 | stride=(head_size, 1), |
| 1411 | assumed_align=4, |
| 1412 | ) |
| 1413 | return ( |
| 1414 | state_cache, |
| 1415 | token_to_req_indices, |
| 1416 | positions, |
| 1417 | slot_mapping, |
| 1418 | block_table, |
| 1419 | compressed_kv, |
| 1420 | ) |
| 1421 | |
| 1422 | @kernel_launcher |
| 1423 | def __call__( |
| 1424 | self, |
| 1425 | *, |
| 1426 | state_cache: torch.Tensor, |
| 1427 | num_actual: int, |
| 1428 | token_to_req_indices: torch.Tensor, |
| 1429 | positions: torch.Tensor, |
| 1430 | slot_mapping: torch.Tensor, |
| 1431 | block_table: torch.Tensor, |
| 1432 | head_dim: int, |
| 1433 | compressed_kv: torch.Tensor | None = None, |
| 1434 | ) -> CuTeDSLLaunchSpec[SparseAttnCompressC128Block8Kernel.CompileKey]: |
| 1435 | compile_key = self.dispatch(head_size=head_dim, state_width=head_dim) |
| 1436 | if compressed_kv is None: |
| 1437 | compressed_kv = torch.empty( |
| 1438 | (num_actual, head_dim), |
| 1439 | dtype=torch.float32, |
| 1440 | device=state_cache.device, |
| 1441 | ) |
| 1442 | launch_args = ( |
| 1443 | state_cache, |
| 1444 | token_to_req_indices, |
| 1445 | positions, |
| 1446 | slot_mapping, |
| 1447 | block_table, |
| 1448 | compressed_kv, |
| 1449 | ) |
| 1450 | return compile_key, launch_args, compressed_kv |
| 1451 | |
| 1452 | |
| 1453 | class SparseAttnNormRopeStoreKernel( |
| 1454 | VllmCuTeDSLJitKernel["SparseAttnNormRopeStoreKernel.CompileKey"] |
| 1455 | ): |
| 1456 | min_scale = 1.0e-4 |
| 1457 | |
| 1458 | @dataclass(frozen=True) |
| 1459 | class CompileKey: |
| 1460 | head_size: int |
| 1461 | rope_head_dim: int |
| 1462 | fp8_max: float |
| 1463 | quant_block: int |
| 1464 | token_stride: int |
| 1465 | scale_dim: int |
| 1466 | kv_block_stride: int |
| 1467 | compress_ratio: int |
| 1468 | norm_weight_dtype: type[cutlass.Numeric] |
| 1469 | kv_cache_block_size: int |
| 1470 | |
| 1471 | @staticmethod |
| 1472 | def kernel(compile_key: CompileKey) -> Any: |
| 1473 | min_scale = SparseAttnNormRopeStoreKernel.min_scale |
| 1474 | head_dim = compile_key.head_size |
| 1475 | rope_dim = compile_key.rope_head_dim |
| 1476 | nope_dim = head_dim - rope_dim |
| 1477 | fp8_max = compile_key.fp8_max |
| 1478 | quant_block = compile_key.quant_block |
| 1479 | token_stride = compile_key.token_stride |
| 1480 | scale_dim = compile_key.scale_dim |
| 1481 | num_warps = head_dim // quant_block |
| 1482 | nope_blocks = nope_dim // quant_block |
| 1483 | tb_size = head_dim // 2 |
| 1484 | compress_ratio = compile_key.compress_ratio |
| 1485 | static_kv_cache_block_size = compile_key.kv_cache_block_size |
| 1486 | |
| 1487 | @cute.kernel |
| 1488 | def device_kernel( |
| 1489 | compressed_kv: cute.Tensor, |
| 1490 | positions: cute.Tensor, |
| 1491 | slot_mapping: cute.Tensor, |
| 1492 | rms_norm_weight: cute.Tensor, |
| 1493 | rms_norm_eps: Float32, |
| 1494 | cos_sin_cache: cute.Tensor, |
| 1495 | k_cache: cute.Tensor, |
| 1496 | kv_slot_mapping: cute.Tensor, |
| 1497 | ): |
| 1498 | token_idx, _, _ = cute.arch.block_idx() |
| 1499 | tid, _, _ = cute.arch.thread_idx() |
| 1500 | warp_id = cute.arch.make_warp_uniform(tid // 32) |
| 1501 | lane_id = tid % 32 |
| 1502 | elem0 = tid * 2 |
| 1503 | |
| 1504 | position = Int64(0) |
| 1505 | kv_slot_idx = Int64(-1) |
| 1506 | has_position = token_idx < positions.shape[0] |
| 1507 | slot_id = Int64(-1) |
| 1508 | if lane_id == 0: |
| 1509 | slot_id = slot_mapping[token_idx] |
| 1510 | if lane_id == 0 and has_position: |
| 1511 | position = positions[token_idx] |
| 1512 | has_kv_slot_idx = token_idx < kv_slot_mapping.shape[0] |
| 1513 | if lane_id == 0 and has_kv_slot_idx: |
| 1514 | kv_slot_idx = kv_slot_mapping[token_idx] |
| 1515 | slot_id = cute.arch.shuffle_sync(slot_id, offset=0) |
| 1516 | position = cute.arch.shuffle_sync(position, offset=0) |
| 1517 | kv_slot_idx = cute.arch.shuffle_sync(kv_slot_idx, offset=0) |
| 1518 | boundary = has_position and ( |
| 1519 | (position + Int64(1)) % Int64(compress_ratio) == Int64(0) |
| 1520 | ) |
| 1521 | active = slot_id >= Int64(0) and boundary and kv_slot_idx >= Int64(0) |
| 1522 | |
| 1523 | if active: |
| 1524 | k_cache_u16 = cute.recast_tensor(k_cache, Uint16) |
| 1525 | k_cache_u32 = cute.recast_tensor(k_cache, Uint32) |
| 1526 | static_block_size = Int64(static_kv_cache_block_size) |
| 1527 | page = kv_slot_idx // static_block_size |
| 1528 | kv_offset = kv_slot_idx - page * static_block_size |
| 1529 | scale_row_offset = static_block_size * Int64(token_stride) |
| 1530 | value_base = page * k_cache.stride[0] + kv_offset * Int64(token_stride) |
| 1531 | scale_base = ( |
| 1532 | page * k_cache.stride[0] |
| 1533 | + scale_row_offset |
| 1534 | + kv_offset * Int64(scale_dim) |
| 1535 | ) |
| 1536 | weight0 = rms_norm_weight[elem0].to(Float32) |
| 1537 | weight1 = rms_norm_weight[elem0 + 1].to(Float32) |
| 1538 | |
| 1539 | base = token_idx.to(Int64) * compressed_kv.stride[0] + elem0.to(Int64) |
| 1540 | x0 = compressed_kv.iterator[base] |
| 1541 | x1 = compressed_kv.iterator[base + Int64(1)] |
| 1542 | |
| 1543 | local_sumsq = x0 * x0 + x1 * x1 |
| 1544 | warp_sum = local_sumsq |
| 1545 | for step in cutlass.range_constexpr(5): |
| 1546 | offset = const_expr(16 >> step) |
| 1547 | warp_sum += cute.arch.shuffle_sync_bfly(warp_sum, offset) |
| 1548 | |
| 1549 | smem = cutlass.utils.SmemAllocator() |
| 1550 | partial_sums = smem.allocate_tensor( |
| 1551 | Float32, cute.make_layout((num_warps,)), byte_alignment=4 |
| 1552 | ) |
| 1553 | |
| 1554 | if lane_id == 0: |
| 1555 | partial_sums[warp_id] = warp_sum |
| 1556 | cute.arch.sync_threads() |
| 1557 | |
| 1558 | total = partial_sums[lane_id % num_warps] |
| 1559 | sum_mask_and_clamp = const_expr( |
| 1560 | (cute.arch.WARP_SIZE - num_warps) << 8 | (cute.arch.WARP_SIZE - 1) |
| 1561 | ) |
| 1562 | for step in cutlass.range_constexpr(3): |
| 1563 | offset = const_expr(4 >> step) |
| 1564 | total += cute.arch.shuffle_sync_bfly( |
| 1565 | total, |
| 1566 | offset, |
| 1567 | mask_and_clamp=sum_mask_and_clamp, |
| 1568 | ) |
| 1569 | |
| 1570 | rrms = cute.math.rsqrt( |
| 1571 | total / Float32(head_dim) + rms_norm_eps, fastmath=True |
| 1572 | ) |
| 1573 | x0 = x0 * rrms * weight0 |
| 1574 | x1 = x1 * rrms * weight1 |
| 1575 | |
| 1576 | if warp_id == nope_blocks: |
| 1577 | pair_idx = lane_id |
| 1578 | compressed_pos = position - Int64(compress_ratio - 1) |
| 1579 | cs_base = compressed_pos * cos_sin_cache.stride[0] + pair_idx.to( |
| 1580 | Int64 |
| 1581 | ) |
| 1582 | cos_v = cos_sin_cache.iterator[cs_base] |
| 1583 | sin_v = cos_sin_cache.iterator[cs_base + Int64(rope_dim // 2)] |
| 1584 | real = x0 * cos_v - x1 * sin_v |
| 1585 | imag = x0 * sin_v + x1 * cos_v |
| 1586 | packed = _fp32x2_to_bf16x2(real, imag) |
| 1587 | out_base = value_base + Int64(nope_dim) + (lane_id * 4).to(Int64) |
| 1588 | k_cache_u32.iterator[out_base // Int64(4)] = packed |
| 1589 | else: |
| 1590 | q_packed = _fp32x2_to_bf16x2(x0, x1) |
| 1591 | q0, q1 = _bf16x2_to_fp32(q_packed) |
| 1592 | abs0 = cute.math.absf(q0) |
| 1593 | abs1 = cute.math.absf(q1) |
| 1594 | local_absmax = cute.arch.fmax(abs0, abs1) |
| 1595 | absmax = local_absmax |
| 1596 | for step in cutlass.range_constexpr(5): |
| 1597 | offset = const_expr(16 >> step) |
| 1598 | absmax = cute.arch.fmax( |
| 1599 | absmax, cute.arch.shuffle_sync_bfly(absmax, offset) |
| 1600 | ) |
| 1601 | scale_raw = cute.arch.fmax( |
| 1602 | Float32(min_scale), |
| 1603 | absmax / Float32(fp8_max), |
| 1604 | ) |
| 1605 | bits = _recast_val(scale_raw, Uint32) |
| 1606 | ue8m0 = ((bits + Uint32(0x7FFFFF)) >> Uint32(23)) & Uint32(0xFF) |
| 1607 | inv_scale = _recast_val( |
| 1608 | (Uint32(254) - ue8m0) << Uint32(23), Float32 |
| 1609 | ) |
| 1610 | y0 = cutlass.min( |
| 1611 | cute.arch.fmax(q0 * inv_scale, Float32(-fp8_max)), |
| 1612 | Float32(fp8_max), |
| 1613 | ) |
| 1614 | y1 = cutlass.min( |
| 1615 | cute.arch.fmax(q1 * inv_scale, Float32(-fp8_max)), |
| 1616 | Float32(fp8_max), |
| 1617 | ) |
| 1618 | packed_fp8 = _fp32x2_to_fp8e4m3x2(y0, y1) |
| 1619 | out_base = value_base + (warp_id * quant_block + lane_id * 2).to( |
| 1620 | Int64 |
| 1621 | ) |
| 1622 | k_cache_u16.iterator[out_base // Int64(2)] = packed_fp8 |
| 1623 | if lane_id == 0: |
| 1624 | k_cache.iterator[scale_base + warp_id.to(Int64)] = ue8m0.to( |
| 1625 | Uint8 |
| 1626 | ) |
| 1627 | if warp_id == 0: |
| 1628 | k_cache.iterator[scale_base + Int64(nope_blocks)] = Uint8(0) |
| 1629 | |
| 1630 | @cute.jit |
| 1631 | def host_entrypoint( |
| 1632 | compressed_kv: cute.Tensor, |
| 1633 | positions: cute.Tensor, |
| 1634 | slot_mapping: cute.Tensor, |
| 1635 | rms_norm_weight: cute.Tensor, |
| 1636 | rms_norm_eps: Float32, |
| 1637 | cos_sin_cache: cute.Tensor, |
| 1638 | k_cache: cute.Tensor, |
| 1639 | kv_slot_mapping: cute.Tensor, |
| 1640 | stream: CUstream, |
| 1641 | ): |
| 1642 | grid = (slot_mapping.shape[0], 1, 1) |
| 1643 | device_kernel( |
| 1644 | compressed_kv, |
| 1645 | positions, |
| 1646 | slot_mapping, |
| 1647 | rms_norm_weight, |
| 1648 | rms_norm_eps, |
| 1649 | cos_sin_cache, |
| 1650 | k_cache, |
| 1651 | kv_slot_mapping, |
| 1652 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 1653 | |
| 1654 | return host_entrypoint |
| 1655 | |
| 1656 | def dispatch( # type: ignore[override] |
| 1657 | self, |
| 1658 | *, |
| 1659 | compress_ratio: int, |
| 1660 | cache_block_size: int, |
| 1661 | cache_alignment: int, |
| 1662 | norm_weight_dtype: type[cutlass.Numeric], |
| 1663 | head_size: int, |
| 1664 | rope_head_dim: int, |
| 1665 | runtime_kv_block_stride: int | None = None, |
| 1666 | ) -> CompileKey: |
| 1667 | raw_kv_cache_block_size = cache_block_size // compress_ratio |
| 1668 | kv_cache_block_size = ( |
| 1669 | raw_kv_cache_block_size if raw_kv_cache_block_size >= 1 else 1 |
| 1670 | ) |
| 1671 | token_stride = 576 |
| 1672 | scale_dim = 8 |
| 1673 | return self.CompileKey( |
| 1674 | head_size=head_size, |
| 1675 | rope_head_dim=rope_head_dim, |
| 1676 | fp8_max=448.0, |
| 1677 | quant_block=64, |
| 1678 | token_stride=token_stride, |
| 1679 | scale_dim=scale_dim, |
| 1680 | kv_block_stride=( |
| 1681 | runtime_kv_block_stride |
| 1682 | if runtime_kv_block_stride is not None |
| 1683 | else round_up( |
| 1684 | kv_cache_block_size * (token_stride + scale_dim), |
| 1685 | cache_alignment, |
| 1686 | ) |
| 1687 | ), |
| 1688 | compress_ratio=compress_ratio, |
| 1689 | norm_weight_dtype=norm_weight_dtype, |
| 1690 | kv_cache_block_size=kv_cache_block_size, |
| 1691 | ) |
| 1692 | |
| 1693 | def get_warmup_keys( |
| 1694 | self, |
| 1695 | vllm_config: Any, |
| 1696 | *, |
| 1697 | k_cache_prefix: str, |
| 1698 | compress_ratio: int, |
| 1699 | ) -> list[CompileKey]: |
| 1700 | hf_config = vllm_config.model_config.hf_config |
| 1701 | cache_config = vllm_config.cache_config |
| 1702 | k_cache_layer = vllm_config.compilation_config.static_forward_context[ |
| 1703 | k_cache_prefix |
| 1704 | ] |
| 1705 | kv_cache = k_cache_layer.kv_cache |
| 1706 | return self._trace_dispatch(self.dispatch)( |
| 1707 | compress_ratio=compress_ratio, |
| 1708 | cache_block_size=cache_config.block_size, |
| 1709 | cache_alignment=576, |
| 1710 | norm_weight_dtype=torch_to_cute_dtype(vllm_config.model_config.dtype), |
| 1711 | head_size=hf_config.head_dim, |
| 1712 | rope_head_dim=hf_config.qk_rope_head_dim, |
| 1713 | runtime_kv_block_stride=kv_cache.stride(0), |
| 1714 | enabled=( |
| 1715 | cache_config.block_size > 0 |
| 1716 | and compress_ratio == 128 |
| 1717 | and kv_cache.ndim == 3 |
| 1718 | and cache_config.cache_dtype == "fp8_ds_mla" |
| 1719 | ), |
| 1720 | _when=lambda *, enabled: enabled, |
| 1721 | ) |
| 1722 | |
| 1723 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 1724 | head_size = compile_key.head_size |
| 1725 | rope_head_dim = compile_key.rope_head_dim |
| 1726 | quant_block = compile_key.quant_block |
| 1727 | token_stride = compile_key.token_stride |
| 1728 | scale_dim = compile_key.scale_dim |
| 1729 | kv_block_stride = compile_key.kv_block_stride |
| 1730 | norm_weight_dtype = compile_key.norm_weight_dtype |
| 1731 | static_kv_cache_block_size = compile_key.kv_cache_block_size |
| 1732 | if quant_block != 64: |
| 1733 | raise ValueError( |
| 1734 | "CuTe DSL sparse-attn store currently requires quant_block=64.", |
| 1735 | ) |
| 1736 | if rope_head_dim != 64: |
| 1737 | raise ValueError( |
| 1738 | "CuTe DSL sparse-attn store currently requires rope_head_dim=64.", |
| 1739 | ) |
| 1740 | if head_size % quant_block != 0: |
| 1741 | raise ValueError("head_size must be divisible by quant_block.") |
| 1742 | if token_stride < head_size + rope_head_dim: |
| 1743 | raise ValueError("token_stride is too small for the packed FP8/BF16 row.") |
| 1744 | expected_scale_dim = (head_size - rope_head_dim) // quant_block + 1 |
| 1745 | if scale_dim < expected_scale_dim: |
| 1746 | raise ValueError("scale_dim is too small for the UE8M0 scale row.") |
| 1747 | if static_kv_cache_block_size <= 0: |
| 1748 | raise ValueError( |
| 1749 | "CuTe DSL sparse-attn store requires a positive static " |
| 1750 | "kv_cache_block_size.", |
| 1751 | ) |
| 1752 | num_positions = cute.sym_int() |
| 1753 | num_slots = cute.sym_int() |
| 1754 | num_kv_slots = cute.sym_int() |
| 1755 | max_pos = cute.sym_int() |
| 1756 | num_blocks = cute.sym_int() |
| 1757 | compressed_kv = cute.runtime.make_fake_tensor( |
| 1758 | Float32, |
| 1759 | (num_slots, head_size), |
| 1760 | stride=(head_size, 1), |
| 1761 | assumed_align=4, |
| 1762 | ) |
| 1763 | positions = make_fake_tensor(Int64, (num_positions,), divisibility=8) |
| 1764 | slot_mapping = make_fake_tensor(Int64, (num_slots,), divisibility=8) |
| 1765 | rms_norm_weight = make_fake_tensor( |
| 1766 | norm_weight_dtype, |
| 1767 | (head_size,), |
| 1768 | divisibility=4, |
| 1769 | ) |
| 1770 | cos_sin_cache = cute.runtime.make_fake_tensor( |
| 1771 | Float32, |
| 1772 | (max_pos, rope_head_dim), |
| 1773 | stride=(rope_head_dim, 1), |
| 1774 | assumed_align=4, |
| 1775 | ) |
| 1776 | k_cache = cute.runtime.make_fake_tensor( |
| 1777 | Uint8, |
| 1778 | (num_blocks, cute.sym_int(), cute.sym_int()), |
| 1779 | stride=(kv_block_stride, cute.sym_int64(divisibility=8), 1), |
| 1780 | assumed_align=16, |
| 1781 | ) |
| 1782 | kv_slot_mapping = make_fake_tensor(Int64, (num_kv_slots,), divisibility=8) |
| 1783 | return ( |
| 1784 | compressed_kv, |
| 1785 | positions, |
| 1786 | slot_mapping, |
| 1787 | rms_norm_weight, |
| 1788 | Float32(0.0), |
| 1789 | cos_sin_cache, |
| 1790 | k_cache, |
| 1791 | kv_slot_mapping, |
| 1792 | ) |
| 1793 | |
| 1794 | @kernel_launcher |
| 1795 | def __call__( |
| 1796 | self, |
| 1797 | *, |
| 1798 | compressed_kv: torch.Tensor, |
| 1799 | positions: torch.Tensor, |
| 1800 | slot_mapping: torch.Tensor, |
| 1801 | rms_norm_weight: torch.Tensor, |
| 1802 | rms_norm_eps: float, |
| 1803 | cos_sin_cache: torch.Tensor, |
| 1804 | kv_cache: torch.Tensor, |
| 1805 | kv_slot_mapping: torch.Tensor, |
| 1806 | compress_ratio: int, |
| 1807 | head_dim: int, |
| 1808 | rope_head_dim: int, |
| 1809 | ) -> CuTeDSLLaunchSpec[SparseAttnNormRopeStoreKernel.CompileKey]: |
| 1810 | compile_key = self.dispatch( |
| 1811 | compress_ratio=compress_ratio, |
| 1812 | cache_block_size=kv_cache.shape[1] * compress_ratio, |
| 1813 | cache_alignment=576, |
| 1814 | norm_weight_dtype=torch_to_cute_dtype(rms_norm_weight.dtype), |
| 1815 | head_size=head_dim, |
| 1816 | rope_head_dim=rope_head_dim, |
| 1817 | runtime_kv_block_stride=kv_cache.stride(0), |
| 1818 | ) |
| 1819 | launch_args = ( |
| 1820 | compressed_kv, |
| 1821 | positions, |
| 1822 | slot_mapping, |
| 1823 | rms_norm_weight, |
| 1824 | rms_norm_eps, |
| 1825 | cos_sin_cache, |
| 1826 | kv_cache, |
| 1827 | kv_slot_mapping, |
| 1828 | ) |
| 1829 | return compile_key, launch_args |
| 1830 | |
| 1831 | |
| 1832 | class SparseAttnNormRopeStoreFullKernel( |
| 1833 | VllmCuTeDSLJitKernel["SparseAttnNormRopeStoreFullKernel.CompileKey"] |
| 1834 | ): |
| 1835 | @dataclass(frozen=True) |
| 1836 | class CompileKey: |
| 1837 | head_size: int |
| 1838 | rope_head_dim: int |
| 1839 | fp8_max: float |
| 1840 | quant_block: int |
| 1841 | compress_ratio: int |
| 1842 | store_full_fp8: bool |
| 1843 | norm_weight_dtype: type[cutlass.Numeric] |
| 1844 | |
| 1845 | @staticmethod |
| 1846 | def kernel(compile_key: CompileKey) -> Any: |
| 1847 | head_dim = compile_key.head_size |
| 1848 | rope_dim = compile_key.rope_head_dim |
| 1849 | nope_dim = head_dim - rope_dim |
| 1850 | fp8_max = compile_key.fp8_max |
| 1851 | quant_block = compile_key.quant_block |
| 1852 | num_warps = head_dim // quant_block |
| 1853 | nope_blocks = nope_dim // quant_block |
| 1854 | tb_size = head_dim // 2 |
| 1855 | compress_ratio = compile_key.compress_ratio |
| 1856 | store_full_fp8 = compile_key.store_full_fp8 |
| 1857 | |
| 1858 | @cute.kernel |
| 1859 | def device_kernel( |
| 1860 | compressed_kv: cute.Tensor, |
| 1861 | positions: cute.Tensor, |
| 1862 | slot_mapping: cute.Tensor, |
| 1863 | rms_norm_weight: cute.Tensor, |
| 1864 | rms_norm_eps: Float32, |
| 1865 | cos_sin_cache: cute.Tensor, |
| 1866 | k_cache: cute.Tensor, |
| 1867 | kv_slot_mapping: cute.Tensor, |
| 1868 | kv_cache_block_size: Int64, |
| 1869 | fp8_scale: cute.Tensor, |
| 1870 | ): |
| 1871 | token_idx, _, _ = cute.arch.block_idx() |
| 1872 | tid, _, _ = cute.arch.thread_idx() |
| 1873 | warp_id = cute.arch.make_warp_uniform(tid // 32) |
| 1874 | lane_id = tid % 32 |
| 1875 | elem0 = tid * 2 |
| 1876 | |
| 1877 | slot_id = slot_mapping[token_idx] |
| 1878 | has_position = token_idx < positions.shape[0] |
| 1879 | position = Int64(0) |
| 1880 | if has_position: |
| 1881 | position = positions[token_idx] |
| 1882 | boundary = has_position and ( |
| 1883 | (position + Int64(1)) % Int64(compress_ratio) == Int64(0) |
| 1884 | ) |
| 1885 | has_kv_slot_idx = token_idx < kv_slot_mapping.shape[0] |
| 1886 | kv_slot_idx = Int64(-1) |
| 1887 | if has_kv_slot_idx: |
| 1888 | kv_slot_idx = kv_slot_mapping[token_idx] |
| 1889 | active = slot_id >= Int64(0) and boundary and kv_slot_idx >= Int64(0) |
| 1890 | |
| 1891 | if active: |
| 1892 | base = token_idx.to(Int64) * compressed_kv.stride[0] + elem0.to(Int64) |
| 1893 | x0 = compressed_kv.iterator[base] |
| 1894 | x1 = compressed_kv.iterator[base + Int64(1)] |
| 1895 | |
| 1896 | local_sumsq = x0 * x0 + x1 * x1 |
| 1897 | warp_sum = local_sumsq |
| 1898 | for step in cutlass.range_constexpr(5): |
| 1899 | offset = const_expr(16 >> step) |
| 1900 | warp_sum += cute.arch.shuffle_sync_bfly(warp_sum, offset) |
| 1901 | |
| 1902 | smem = cutlass.utils.SmemAllocator() |
| 1903 | partial_sums = smem.allocate_tensor( |
| 1904 | Float32, cute.make_layout((num_warps,)), byte_alignment=4 |
| 1905 | ) |
| 1906 | rrms_shared = smem.allocate_tensor( |
| 1907 | Float32, cute.make_layout((1,)), byte_alignment=4 |
| 1908 | ) |
| 1909 | |
| 1910 | if lane_id == 0: |
| 1911 | partial_sums[warp_id] = warp_sum |
| 1912 | cute.arch.sync_threads() |
| 1913 | if tid == 0: |
| 1914 | total = Float32(0.0) |
| 1915 | for i in cutlass.range_constexpr(num_warps): |
| 1916 | total += partial_sums[i] |
| 1917 | rrms_shared[0] = cute.math.rsqrt( |
| 1918 | total / Float32(head_dim) + rms_norm_eps, fastmath=True |
| 1919 | ) |
| 1920 | cute.arch.sync_threads() |
| 1921 | |
| 1922 | rrms = rrms_shared[0] |
| 1923 | x0 = x0 * rrms * rms_norm_weight[elem0].to(Float32) |
| 1924 | x1 = x1 * rrms * rms_norm_weight[elem0 + 1].to(Float32) |
| 1925 | |
| 1926 | page = kv_slot_idx // kv_cache_block_size |
| 1927 | kv_offset = kv_slot_idx - page * kv_cache_block_size |
| 1928 | value_base = page * k_cache.stride[0] + kv_offset * k_cache.stride[1] |
| 1929 | |
| 1930 | if const_expr(store_full_fp8): |
| 1931 | k_cache_u16 = cute.recast_tensor(k_cache, Uint16) |
| 1932 | inv_fp8 = Float32(1.0) / fp8_scale[0] |
| 1933 | fp8_v0 = x0 |
| 1934 | fp8_v1 = x1 |
| 1935 | if warp_id == nope_blocks: |
| 1936 | compressed_pos = (position // Int64(compress_ratio)) * Int64( |
| 1937 | compress_ratio |
| 1938 | ) |
| 1939 | pair_idx = lane_id |
| 1940 | cs_base = compressed_pos * cos_sin_cache.stride[ |
| 1941 | 0 |
| 1942 | ] + pair_idx.to(Int64) |
| 1943 | cos_v = cos_sin_cache.iterator[cs_base] |
| 1944 | sin_v = cos_sin_cache.iterator[cs_base + Int64(rope_dim // 2)] |
| 1945 | fp8_v0 = x0 * cos_v - x1 * sin_v |
| 1946 | fp8_v1 = x0 * sin_v + x1 * cos_v |
| 1947 | fp8_packed_bf16 = _fp32x2_to_bf16x2(fp8_v0, fp8_v1) |
| 1948 | b0, b1 = _bf16x2_to_fp32(fp8_packed_bf16) |
| 1949 | y0 = cutlass.min( |
| 1950 | cutlass.max(b0 * inv_fp8, Float32(-fp8_max)), |
| 1951 | Float32(fp8_max), |
| 1952 | ) |
| 1953 | y1 = cutlass.min( |
| 1954 | cutlass.max(b1 * inv_fp8, Float32(-fp8_max)), |
| 1955 | Float32(fp8_max), |
| 1956 | ) |
| 1957 | packed_fp8 = _fp32x2_to_fp8e4m3x2(y0, y1) |
| 1958 | out_base = value_base + elem0.to(Int64) |
| 1959 | k_cache_u16.iterator[out_base // Int64(2)] = packed_fp8 |
| 1960 | else: |
| 1961 | k_cache_u32 = cute.recast_tensor(k_cache, Uint32) |
| 1962 | bf16_v0 = x0 |
| 1963 | bf16_v1 = x1 |
| 1964 | if warp_id == nope_blocks: |
| 1965 | compressed_pos = (position // Int64(compress_ratio)) * Int64( |
| 1966 | compress_ratio |
| 1967 | ) |
| 1968 | pair_idx = lane_id |
| 1969 | cs_base = compressed_pos * cos_sin_cache.stride[ |
| 1970 | 0 |
| 1971 | ] + pair_idx.to(Int64) |
| 1972 | cos_v = cos_sin_cache.iterator[cs_base] |
| 1973 | sin_v = cos_sin_cache.iterator[cs_base + Int64(rope_dim // 2)] |
| 1974 | bf16_v0 = x0 * cos_v - x1 * sin_v |
| 1975 | bf16_v1 = x0 * sin_v + x1 * cos_v |
| 1976 | bf16_packed = _fp32x2_to_bf16x2(bf16_v0, bf16_v1) |
| 1977 | out_base = value_base + (elem0 * 2).to(Int64) |
| 1978 | k_cache_u32.iterator[out_base // Int64(4)] = bf16_packed |
| 1979 | |
| 1980 | @cute.jit |
| 1981 | def host_entrypoint( |
| 1982 | compressed_kv: cute.Tensor, |
| 1983 | positions: cute.Tensor, |
| 1984 | slot_mapping: cute.Tensor, |
| 1985 | rms_norm_weight: cute.Tensor, |
| 1986 | rms_norm_eps: Float32, |
| 1987 | cos_sin_cache: cute.Tensor, |
| 1988 | k_cache: cute.Tensor, |
| 1989 | kv_slot_mapping: cute.Tensor, |
| 1990 | kv_cache_block_size: Int64, |
| 1991 | fp8_scale: cute.Tensor, |
| 1992 | stream: CUstream, |
| 1993 | ): |
| 1994 | grid = (slot_mapping.shape[0], 1, 1) |
| 1995 | device_kernel( |
| 1996 | compressed_kv, |
| 1997 | positions, |
| 1998 | slot_mapping, |
| 1999 | rms_norm_weight, |
| 2000 | rms_norm_eps, |
| 2001 | cos_sin_cache, |
| 2002 | k_cache, |
| 2003 | kv_slot_mapping, |
| 2004 | kv_cache_block_size, |
| 2005 | fp8_scale, |
| 2006 | ).launch(grid=grid, block=(tb_size, 1, 1), stream=stream) |
| 2007 | |
| 2008 | return host_entrypoint |
| 2009 | |
| 2010 | def dispatch( # type: ignore[override] |
| 2011 | self, |
| 2012 | *, |
| 2013 | store_full_fp8: bool, |
| 2014 | norm_weight_dtype: type[cutlass.Numeric], |
| 2015 | head_size: int, |
| 2016 | rope_head_dim: int, |
| 2017 | compress_ratio: int, |
| 2018 | ) -> CompileKey: |
| 2019 | return self.CompileKey( |
| 2020 | head_size=head_size, |
| 2021 | rope_head_dim=rope_head_dim, |
| 2022 | compress_ratio=compress_ratio, |
| 2023 | fp8_max=448.0, |
| 2024 | quant_block=64, |
| 2025 | store_full_fp8=store_full_fp8, |
| 2026 | norm_weight_dtype=norm_weight_dtype, |
| 2027 | ) |
| 2028 | |
| 2029 | def get_warmup_keys(self, vllm_config: Any) -> list[CompileKey]: |
| 2030 | hf_config = vllm_config.model_config.hf_config |
| 2031 | cache_config = vllm_config.cache_config |
| 2032 | return self._trace_dispatch(self.dispatch)( |
| 2033 | compress_ratio=128, |
| 2034 | store_full_fp8=( |
| 2035 | cache_config.cache_dtype.startswith("fp8") |
| 2036 | and cache_config.cache_dtype != "fp8_ds_mla" |
| 2037 | ), |
| 2038 | norm_weight_dtype=torch_to_cute_dtype(vllm_config.model_config.dtype), |
| 2039 | head_size=hf_config.head_dim, |
| 2040 | rope_head_dim=hf_config.qk_rope_head_dim, |
| 2041 | enabled=( |
| 2042 | cache_config.block_size > 0 |
| 2043 | and 128 in hf_config.compress_ratios |
| 2044 | and cache_config.cache_dtype != "fp8_ds_mla" |
| 2045 | ), |
| 2046 | _when=lambda *, enabled: enabled, |
| 2047 | ) |
| 2048 | |
| 2049 | def warmup_inputs(self, compile_key: CompileKey) -> tuple[Any, ...]: |
| 2050 | head_size = compile_key.head_size |
| 2051 | rope_head_dim = compile_key.rope_head_dim |
| 2052 | quant_block = compile_key.quant_block |
| 2053 | norm_weight_dtype = compile_key.norm_weight_dtype |
| 2054 | if quant_block != 64: |
| 2055 | raise ValueError( |
| 2056 | "CuTe DSL sparse-attn store currently requires quant_block=64.", |
| 2057 | ) |
| 2058 | if rope_head_dim != 64: |
| 2059 | raise ValueError( |
| 2060 | "CuTe DSL sparse-attn store currently requires rope_head_dim=64.", |
| 2061 | ) |
| 2062 | if head_size % quant_block != 0: |
| 2063 | raise ValueError("head_size must be divisible by quant_block.") |
| 2064 | num_positions = cute.sym_int() |
| 2065 | num_slots = cute.sym_int() |
| 2066 | num_kv_slots = cute.sym_int() |
| 2067 | max_pos = cute.sym_int() |
| 2068 | num_blocks = cute.sym_int() |
| 2069 | compressed_kv = cute.runtime.make_fake_tensor( |
| 2070 | Float32, |
| 2071 | (num_slots, head_size), |
| 2072 | stride=(cute.sym_int64(divisibility=4), 1), |
| 2073 | assumed_align=4, |
| 2074 | ) |
| 2075 | positions = make_fake_tensor(Int64, (num_positions,), divisibility=8) |
| 2076 | slot_mapping = make_fake_tensor(Int64, (num_slots,), divisibility=8) |
| 2077 | rms_norm_weight = make_fake_tensor( |
| 2078 | norm_weight_dtype, |
| 2079 | (head_size,), |
| 2080 | divisibility=4, |
| 2081 | ) |
| 2082 | cos_sin_cache = cute.runtime.make_fake_tensor( |
| 2083 | Float32, |
| 2084 | (max_pos, rope_head_dim), |
| 2085 | stride=(cute.sym_int64(divisibility=4), 1), |
| 2086 | assumed_align=4, |
| 2087 | ) |
| 2088 | k_cache = cute.runtime.make_fake_tensor( |
| 2089 | Uint8, |
| 2090 | (num_blocks, cute.sym_int(), cute.sym_int()), |
| 2091 | stride=(cute.sym_int64(divisibility=16), cute.sym_int64(divisibility=8), 1), |
| 2092 | assumed_align=16, |
| 2093 | ) |
| 2094 | kv_slot_mapping = make_fake_tensor(Int64, (num_kv_slots,), divisibility=8) |
| 2095 | fp8_scale = make_fake_tensor(Float32, (1,), divisibility=1) |
| 2096 | return ( |
| 2097 | compressed_kv, |
| 2098 | positions, |
| 2099 | slot_mapping, |
| 2100 | rms_norm_weight, |
| 2101 | Float32(0.0), |
| 2102 | cos_sin_cache, |
| 2103 | k_cache, |
| 2104 | kv_slot_mapping, |
| 2105 | Int64(0), |
| 2106 | fp8_scale, |
| 2107 | ) |
| 2108 | |
| 2109 | @kernel_launcher |
| 2110 | def __call__( |
| 2111 | self, |
| 2112 | *, |
| 2113 | compressed_kv: torch.Tensor, |
| 2114 | positions: torch.Tensor, |
| 2115 | slot_mapping: torch.Tensor, |
| 2116 | rms_norm_weight: torch.Tensor, |
| 2117 | rms_norm_eps: float, |
| 2118 | cos_sin_cache: torch.Tensor, |
| 2119 | kv_cache: torch.Tensor, |
| 2120 | kv_slot_mapping: torch.Tensor, |
| 2121 | compress_ratio: int, |
| 2122 | store_full_fp8: bool, |
| 2123 | fp8_scale: torch.Tensor | None, |
| 2124 | head_dim: int, |
| 2125 | rope_head_dim: int, |
| 2126 | ) -> CuTeDSLLaunchSpec[SparseAttnNormRopeStoreFullKernel.CompileKey]: |
| 2127 | compile_key = self.dispatch( |
| 2128 | compress_ratio=compress_ratio, |
| 2129 | store_full_fp8=store_full_fp8, |
| 2130 | norm_weight_dtype=torch_to_cute_dtype(rms_norm_weight.dtype), |
| 2131 | head_size=head_dim, |
| 2132 | rope_head_dim=rope_head_dim, |
| 2133 | ) |
| 2134 | if fp8_scale is None: |
| 2135 | fp8_scale = torch.ones(1, dtype=torch.float32, device=kv_cache.device) |
| 2136 | launch_args = ( |
| 2137 | compressed_kv, |
| 2138 | positions, |
| 2139 | slot_mapping, |
| 2140 | rms_norm_weight, |
| 2141 | rms_norm_eps, |
| 2142 | cos_sin_cache, |
| 2143 | kv_cache.view(torch.uint8), |
| 2144 | kv_slot_mapping, |
| 2145 | kv_cache.shape[1], |
| 2146 | fp8_scale, |
| 2147 | ) |
| 2148 | return compile_key, launch_args |
| 2149 | |
| 2150 | |
| 2151 | def fused_kv_compress_norm_rope_insert_sparse_attn_cutedsl( |
| 2152 | state_cache: torch.Tensor, |
| 2153 | token_to_req_indices: torch.Tensor, |
| 2154 | positions: torch.Tensor, |
| 2155 | slot_mapping: torch.Tensor, |
| 2156 | block_table: torch.Tensor, |
| 2157 | block_size: int, |
| 2158 | rms_norm_weight: torch.Tensor, |
| 2159 | rms_norm_eps: float, |
| 2160 | cos_sin_cache: torch.Tensor, |
| 2161 | k_cache: torch.Tensor, |
| 2162 | kv_slot_mapping: torch.Tensor, |
| 2163 | kv_cache_block_size: int, |
| 2164 | kv_block_stride: int, |
| 2165 | head_size: int = 512, |
| 2166 | state_width: int = 1024, |
| 2167 | rope_head_dim: int = 64, |
| 2168 | fp8_max: float = 448.0, |
| 2169 | quant_block: int = 64, |
| 2170 | token_stride: int = 576, |
| 2171 | scale_dim: int = 8, |
| 2172 | compress_ratio: int = 4, |
| 2173 | overlap: bool = True, |
| 2174 | store_full_kv: bool = False, |
| 2175 | store_full_fp8: bool = False, |
| 2176 | fp8_scale: torch.Tensor | None = None, |
| 2177 | ) -> None: |
| 2178 | del fp8_max, quant_block, token_stride, scale_dim, kv_block_stride |
| 2179 | if not ( |
| 2180 | head_size == 512 |
| 2181 | and state_width == 2 * head_size |
| 2182 | and compress_ratio == 4 |
| 2183 | and overlap |
| 2184 | ): |
| 2185 | raise ValueError( |
| 2186 | "CuTe DSL fused sparse-attn wrapper only supports the real " |
| 2187 | "DeepSeek V4 C4 layout: head_size=512, state_width=1024, " |
| 2188 | "compress_ratio=4, overlap=True." |
| 2189 | ) |
| 2190 | if k_cache.ndim != 3: |
| 2191 | raise ValueError( |
| 2192 | "CuTe DSL sparse-attn fused store expects the real DeepSeek V4 " |
| 2193 | f"3D k_cache layout [num_blocks, block_size, 584], got ndim={k_cache.ndim}." |
| 2194 | ) |
| 2195 | if kv_cache_block_size != k_cache.shape[1]: |
| 2196 | raise ValueError( |
| 2197 | "CuTe DSL fused sparse-attn wrapper expected kv_cache_block_size " |
| 2198 | f"to match k_cache.shape[1], got {kv_cache_block_size} and " |
| 2199 | f"{k_cache.shape[1]}." |
| 2200 | ) |
| 2201 | if positions.numel() == 0: |
| 2202 | return |
| 2203 | if store_full_fp8 and not store_full_kv: |
| 2204 | raise ValueError("store_full_fp8 requires store_full_kv.") |
| 2205 | |
| 2206 | if store_full_kv: |
| 2207 | _SPARSE_ATTN_COMPRESS_NORM_ROPE_STORE_FULL_C4_KERNEL( |
| 2208 | state_cache=state_cache, |
| 2209 | token_to_req_indices=token_to_req_indices, |
| 2210 | positions=positions, |
| 2211 | slot_mapping=slot_mapping, |
| 2212 | block_table=block_table, |
| 2213 | block_size=block_size, |
| 2214 | rms_norm_weight=rms_norm_weight, |
| 2215 | rms_norm_eps=rms_norm_eps, |
| 2216 | cos_sin_cache=cos_sin_cache, |
| 2217 | kv_cache=k_cache, |
| 2218 | kv_slot_mapping=kv_slot_mapping, |
| 2219 | compress_ratio=compress_ratio, |
| 2220 | store_full_fp8=store_full_fp8, |
| 2221 | fp8_scale=fp8_scale, |
| 2222 | head_dim=head_size, |
| 2223 | rope_head_dim=rope_head_dim, |
| 2224 | ) |
| 2225 | return |
| 2226 | _SPARSE_ATTN_COMPRESS_NORM_ROPE_STORE_C4_KERNEL( |
| 2227 | state_cache=state_cache, |
| 2228 | token_to_req_indices=token_to_req_indices, |
| 2229 | positions=positions, |
| 2230 | slot_mapping=slot_mapping, |
| 2231 | block_table=block_table, |
| 2232 | block_size=block_size, |
| 2233 | rms_norm_weight=rms_norm_weight, |
| 2234 | rms_norm_eps=rms_norm_eps, |
| 2235 | cos_sin_cache=cos_sin_cache, |
| 2236 | kv_cache=k_cache, |
| 2237 | kv_slot_mapping=kv_slot_mapping, |
| 2238 | compress_ratio=compress_ratio, |
| 2239 | head_dim=head_size, |
| 2240 | rope_head_dim=rope_head_dim, |
| 2241 | ) |
| 2242 | |
| 2243 | |
| 2244 | def split_kv_compress_norm_rope_insert_sparse_attn_cutedsl( |
| 2245 | state_cache: torch.Tensor, |
| 2246 | token_to_req_indices: torch.Tensor, |
| 2247 | positions: torch.Tensor, |
| 2248 | slot_mapping: torch.Tensor, |
| 2249 | block_table: torch.Tensor, |
| 2250 | block_size: int, |
| 2251 | compressed_kv: torch.Tensor, |
| 2252 | rms_norm_weight: torch.Tensor, |
| 2253 | rms_norm_eps: float, |
| 2254 | cos_sin_cache: torch.Tensor, |
| 2255 | k_cache: torch.Tensor, |
| 2256 | kv_slot_mapping: torch.Tensor, |
| 2257 | kv_cache_block_size: int, |
| 2258 | kv_block_stride: int, |
| 2259 | head_size: int = 512, |
| 2260 | state_width: int = 512, |
| 2261 | rope_head_dim: int = 64, |
| 2262 | fp8_max: float = 448.0, |
| 2263 | quant_block: int = 64, |
| 2264 | token_stride: int = 576, |
| 2265 | scale_dim: int = 8, |
| 2266 | compress_ratio: int = 128, |
| 2267 | overlap: bool = False, |
| 2268 | store_full_kv: bool = False, |
| 2269 | store_full_fp8: bool = False, |
| 2270 | fp8_scale: torch.Tensor | None = None, |
| 2271 | ) -> None: |
| 2272 | del fp8_max, quant_block, token_stride, scale_dim, kv_block_stride |
| 2273 | if not ( |
| 2274 | head_size == 512 |
| 2275 | and state_width == head_size |
| 2276 | and compress_ratio == 128 |
| 2277 | and not overlap |
| 2278 | and block_size == 8 |
| 2279 | ): |
| 2280 | raise ValueError( |
| 2281 | "CuTe DSL split sparse-attn wrapper only supports the real " |
| 2282 | "DeepSeek V4 C128 layout: head_size=512, state_width=512, " |
| 2283 | "compress_ratio=128, overlap=False, block_size=8." |
| 2284 | ) |
| 2285 | if k_cache.ndim != 3: |
| 2286 | raise ValueError( |
| 2287 | "CuTe DSL sparse-attn store expects the real DeepSeek V4 " |
| 2288 | f"3D k_cache layout [num_blocks, block_size, 584], got ndim={k_cache.ndim}." |
| 2289 | ) |
| 2290 | if not store_full_kv and kv_cache_block_size != k_cache.shape[1]: |
| 2291 | raise ValueError( |
| 2292 | "CuTe DSL split sparse-attn wrapper expected kv_cache_block_size " |
| 2293 | f"to match k_cache.shape[1], got {kv_cache_block_size} and " |
| 2294 | f"{k_cache.shape[1]}." |
| 2295 | ) |
| 2296 | if positions.numel() == 0: |
| 2297 | return |
| 2298 | if store_full_fp8 and not store_full_kv: |
| 2299 | raise ValueError("store_full_fp8 requires store_full_kv.") |
| 2300 | |
| 2301 | _SPARSE_ATTN_COMPRESS_C128_BLOCK8_KERNEL( |
| 2302 | state_cache=state_cache, |
| 2303 | num_actual=compressed_kv.shape[0], |
| 2304 | token_to_req_indices=token_to_req_indices, |
| 2305 | positions=positions, |
| 2306 | slot_mapping=slot_mapping, |
| 2307 | block_table=block_table, |
| 2308 | head_dim=head_size, |
| 2309 | compressed_kv=compressed_kv, |
| 2310 | ) |
| 2311 | if store_full_kv: |
| 2312 | _SPARSE_ATTN_NORM_ROPE_STORE_FULL_KERNEL( |
| 2313 | compressed_kv=compressed_kv, |
| 2314 | positions=positions, |
| 2315 | slot_mapping=slot_mapping, |
| 2316 | rms_norm_weight=rms_norm_weight, |
| 2317 | rms_norm_eps=rms_norm_eps, |
| 2318 | cos_sin_cache=cos_sin_cache, |
| 2319 | kv_cache=k_cache, |
| 2320 | kv_slot_mapping=kv_slot_mapping, |
| 2321 | compress_ratio=compress_ratio, |
| 2322 | store_full_fp8=store_full_fp8, |
| 2323 | fp8_scale=fp8_scale, |
| 2324 | head_dim=head_size, |
| 2325 | rope_head_dim=rope_head_dim, |
| 2326 | ) |
| 2327 | return |
| 2328 | _SPARSE_ATTN_NORM_ROPE_STORE_KERNEL( |
| 2329 | compressed_kv=compressed_kv, |
| 2330 | positions=positions, |
| 2331 | slot_mapping=slot_mapping, |
| 2332 | rms_norm_weight=rms_norm_weight, |
| 2333 | rms_norm_eps=rms_norm_eps, |
| 2334 | cos_sin_cache=cos_sin_cache, |
| 2335 | kv_cache=k_cache, |
| 2336 | kv_slot_mapping=kv_slot_mapping, |
| 2337 | compress_ratio=compress_ratio, |
| 2338 | head_dim=head_size, |
| 2339 | rope_head_dim=rope_head_dim, |
| 2340 | ) |
| 2341 | |
| 2342 | |
| 2343 | class SparseAttnCompressorCuteDSL: |
| 2344 | def __init__( |
| 2345 | self, |
| 2346 | *, |
| 2347 | c4_store: SparseAttnCompressNormRopeStoreC4Kernel, |
| 2348 | c4_store_full: SparseAttnCompressNormRopeStoreFullC4Kernel, |
| 2349 | c128_compress: SparseAttnCompressC128Block8Kernel, |
| 2350 | c128_store: SparseAttnNormRopeStoreKernel, |
| 2351 | c128_store_full: SparseAttnNormRopeStoreFullKernel, |
| 2352 | ) -> None: |
| 2353 | self.c4_store = c4_store |
| 2354 | self.c4_store_full = c4_store_full |
| 2355 | self.c128_compress = c128_compress |
| 2356 | self.c128_store = c128_store |
| 2357 | self.c128_store_full = c128_store_full |
| 2358 | |
| 2359 | def __call__( |
| 2360 | self, |
| 2361 | state_cache: torch.Tensor, |
| 2362 | num_actual: int, |
| 2363 | token_to_req_indices: torch.Tensor, |
| 2364 | positions: torch.Tensor, |
| 2365 | slot_mapping: torch.Tensor, |
| 2366 | block_table: torch.Tensor, |
| 2367 | block_size: int, |
| 2368 | state_width: int, |
| 2369 | cos_sin_cache: torch.Tensor, |
| 2370 | kv_cache: torch.Tensor, |
| 2371 | k_cache_metadata: Any, |
| 2372 | pdl_kwargs: dict, |
| 2373 | head_dim: int, |
| 2374 | rope_head_dim: int, |
| 2375 | compress_ratio: int, |
| 2376 | overlap: bool, |
| 2377 | use_fp4_cache: bool, |
| 2378 | rms_norm_weight: torch.Tensor, |
| 2379 | rms_norm_eps: float, |
| 2380 | quant_block: int, |
| 2381 | token_stride: int, |
| 2382 | scale_dim: int, |
| 2383 | store_full_kv: bool = False, |
| 2384 | store_full_fp8: bool = False, |
| 2385 | fp8_scale: torch.Tensor | None = None, |
| 2386 | ) -> None: |
| 2387 | del pdl_kwargs, state_width, overlap, use_fp4_cache |
| 2388 | del quant_block, token_stride, scale_dim |
| 2389 | if positions.numel() == 0: |
| 2390 | return |
| 2391 | if kv_cache.ndim != 3: |
| 2392 | raise ValueError( |
| 2393 | "CuTe DSL sparse-attn compressor expects the real DeepSeek V4 " |
| 2394 | f"3D kv_cache layout, got ndim={kv_cache.ndim}.", |
| 2395 | ) |
| 2396 | if store_full_fp8 and (not store_full_kv): |
| 2397 | raise ValueError("store_full_fp8 requires store_full_kv.") |
| 2398 | if compress_ratio == 4: |
| 2399 | if store_full_kv: |
| 2400 | self.c4_store_full( |
| 2401 | state_cache=state_cache, |
| 2402 | token_to_req_indices=token_to_req_indices, |
| 2403 | positions=positions, |
| 2404 | slot_mapping=slot_mapping, |
| 2405 | block_table=block_table, |
| 2406 | block_size=block_size, |
| 2407 | rms_norm_weight=rms_norm_weight, |
| 2408 | rms_norm_eps=rms_norm_eps, |
| 2409 | cos_sin_cache=cos_sin_cache, |
| 2410 | kv_cache=kv_cache, |
| 2411 | kv_slot_mapping=k_cache_metadata.slot_mapping, |
| 2412 | compress_ratio=compress_ratio, |
| 2413 | store_full_fp8=store_full_fp8, |
| 2414 | fp8_scale=fp8_scale, |
| 2415 | head_dim=head_dim, |
| 2416 | rope_head_dim=rope_head_dim, |
| 2417 | ) |
| 2418 | else: |
| 2419 | self.c4_store( |
| 2420 | state_cache=state_cache, |
| 2421 | token_to_req_indices=token_to_req_indices, |
| 2422 | positions=positions, |
| 2423 | slot_mapping=slot_mapping, |
| 2424 | block_table=block_table, |
| 2425 | block_size=block_size, |
| 2426 | rms_norm_weight=rms_norm_weight, |
| 2427 | rms_norm_eps=rms_norm_eps, |
| 2428 | cos_sin_cache=cos_sin_cache, |
| 2429 | kv_cache=kv_cache, |
| 2430 | kv_slot_mapping=k_cache_metadata.slot_mapping, |
| 2431 | compress_ratio=compress_ratio, |
| 2432 | head_dim=head_dim, |
| 2433 | rope_head_dim=rope_head_dim, |
| 2434 | ) |
| 2435 | return |
| 2436 | if compress_ratio != 128: |
| 2437 | raise ValueError( |
| 2438 | "CuTe DSL sparse-attn compressor supports compress_ratio 4 or " |
| 2439 | f"128, got {compress_ratio}.", |
| 2440 | ) |
| 2441 | if block_size != self.c128_compress.state_block_size: |
| 2442 | raise ValueError( |
| 2443 | "CuTe DSL C128 sparse-attn compressor requires state-cache " |
| 2444 | f"block_size={self.c128_compress.state_block_size}, got " |
| 2445 | f"{block_size}." |
| 2446 | ) |
| 2447 | compressed_kv = self.c128_compress( |
| 2448 | state_cache=state_cache, |
| 2449 | num_actual=num_actual, |
| 2450 | token_to_req_indices=token_to_req_indices, |
| 2451 | positions=positions, |
| 2452 | slot_mapping=slot_mapping, |
| 2453 | block_table=block_table, |
| 2454 | head_dim=head_dim, |
| 2455 | ) |
| 2456 | if store_full_kv: |
| 2457 | self.c128_store_full( |
| 2458 | compressed_kv=compressed_kv, |
| 2459 | positions=positions, |
| 2460 | slot_mapping=slot_mapping, |
| 2461 | rms_norm_weight=rms_norm_weight, |
| 2462 | rms_norm_eps=rms_norm_eps, |
| 2463 | cos_sin_cache=cos_sin_cache, |
| 2464 | kv_cache=kv_cache, |
| 2465 | kv_slot_mapping=k_cache_metadata.slot_mapping, |
| 2466 | compress_ratio=compress_ratio, |
| 2467 | store_full_fp8=store_full_fp8, |
| 2468 | fp8_scale=fp8_scale, |
| 2469 | head_dim=head_dim, |
| 2470 | rope_head_dim=rope_head_dim, |
| 2471 | ) |
| 2472 | return |
| 2473 | self.c128_store( |
| 2474 | compressed_kv=compressed_kv, |
| 2475 | positions=positions, |
| 2476 | slot_mapping=slot_mapping, |
| 2477 | rms_norm_weight=rms_norm_weight, |
| 2478 | rms_norm_eps=rms_norm_eps, |
| 2479 | cos_sin_cache=cos_sin_cache, |
| 2480 | kv_cache=kv_cache, |
| 2481 | kv_slot_mapping=k_cache_metadata.slot_mapping, |
| 2482 | compress_ratio=compress_ratio, |
| 2483 | head_dim=head_dim, |
| 2484 | rope_head_dim=rope_head_dim, |
| 2485 | ) |
| 2486 | |
| 2487 | |
| 2488 | _SPARSE_ATTN_COMPRESS_NORM_ROPE_STORE_C4_KERNEL = ( |
| 2489 | SparseAttnCompressNormRopeStoreC4Kernel() |
| 2490 | ) |
| 2491 | _SPARSE_ATTN_COMPRESS_NORM_ROPE_STORE_FULL_C4_KERNEL = ( |
| 2492 | SparseAttnCompressNormRopeStoreFullC4Kernel() |
| 2493 | ) |
| 2494 | _SPARSE_ATTN_COMPRESS_C128_BLOCK8_KERNEL = SparseAttnCompressC128Block8Kernel() |
| 2495 | _SPARSE_ATTN_NORM_ROPE_STORE_KERNEL = SparseAttnNormRopeStoreKernel() |
| 2496 | _SPARSE_ATTN_NORM_ROPE_STORE_FULL_KERNEL = SparseAttnNormRopeStoreFullKernel() |
| 2497 | _SPARSE_ATTN_COMPRESSOR_CUTEDSL_KERNEL = SparseAttnCompressorCuteDSL( |
| 2498 | c4_store=_SPARSE_ATTN_COMPRESS_NORM_ROPE_STORE_C4_KERNEL, |
| 2499 | c4_store_full=_SPARSE_ATTN_COMPRESS_NORM_ROPE_STORE_FULL_C4_KERNEL, |
| 2500 | c128_compress=_SPARSE_ATTN_COMPRESS_C128_BLOCK8_KERNEL, |
| 2501 | c128_store=_SPARSE_ATTN_NORM_ROPE_STORE_KERNEL, |
| 2502 | c128_store_full=_SPARSE_ATTN_NORM_ROPE_STORE_FULL_KERNEL, |
| 2503 | ) |