| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """CPU-kernel-vs-eager-reference parity tests for the DeepSeek-V4 CPU port's |
| 4 | sgl-kernels ops (csrc/cpu/sgl-kernels/). Each test carries its own reference |
| 5 | reimplementation of the kernel's math rather than depending on production |
| 6 | model code. |
| 7 | """ |
| 8 | |
| 9 | from unittest import mock |
| 10 | |
| 11 | import pytest |
| 12 | import torch |
| 13 | |
| 14 | import vllm.model_executor.kernels.mhc as mhc_kernels |
| 15 | from tests.kernels.test_fused_deepseek_v4_qnorm_rope_kv_insert import ( |
| 16 | HEAD_DIM, |
| 17 | NOPE_DIM, |
| 18 | QUANT_BLOCK, |
| 19 | ROPE_DIM, |
| 20 | make_cos_sin_cache, |
| 21 | ) |
| 22 | from tests.kernels.utils import bf16_ulp_distance, fp8_ulp_distance |
| 23 | from vllm._custom_ops import ( |
| 24 | compress_norm_rope_store_cpu, |
| 25 | compress_norm_rope_store_indexer_cpu, |
| 26 | flash_mla_with_kvcache_cpu, |
| 27 | fp8_paged_mqa_logits_cpu, |
| 28 | fused_indexer_q_rope_quant_cpu, |
| 29 | fused_qnorm_rope_kv_insert_cpu, |
| 30 | inverse_gptj_rope_o_proj_cpu, |
| 31 | save_partial_states_cpu, |
| 32 | ) |
| 33 | from vllm.models.deepseek_v4.common.ops import fused_indexer_q_rope_quant |
| 34 | from vllm.platforms import current_platform |
| 35 | from vllm.utils.torch_utils import set_random_seed |
| 36 | |
| 37 | DEVICE = current_platform.device_type |
| 38 | |
| 39 | # Main attention cache (fp8_ds_mla): NoPE fp8 in QUANT_BLOCK-wide UE8M0 blocks, |
| 40 | # RoPE bf16, one scale byte per block (+1 pad byte) per token. |
| 41 | SCALE_BYTES_PER_TOKEN = NOPE_DIM // QUANT_BLOCK + 1 |
| 42 | TOKEN_DATA_BYTES = NOPE_DIM + ROPE_DIM * 2 |
| 43 | FP8_MAX = 448.0 |
| 44 | |
| 45 | # Indexer K-cache: all 128 dims in one fp8 block, one raw fp32 scale per token. |
| 46 | INDEXER_HEAD_DIM = 128 |
| 47 | INDEXER_NOPE_DIM = INDEXER_HEAD_DIM - ROPE_DIM |
| 48 | INDEXER_SCALE_BYTES = 4 |
| 49 | |
| 50 | |
| 51 | def _op_available(*names: str) -> bool: |
| 52 | return all(hasattr(torch.ops._C, name) for name in names) |
| 53 | |
| 54 | |
| 55 | pytestmark = pytest.mark.skipif( |
| 56 | not current_platform.is_cpu() |
| 57 | or not _op_available( |
| 58 | "save_partial_states_cpu", |
| 59 | "compress_norm_rope_store_cpu", |
| 60 | "compress_norm_rope_store_indexer_cpu", |
| 61 | "flash_mla_with_kvcache_cpu", |
| 62 | "fused_qnorm_rope_kv_insert_cpu", |
| 63 | "inverse_gptj_rope_o_proj_cpu", |
| 64 | "fp8_paged_mqa_logits_cpu", |
| 65 | "fused_indexer_q_rope_quant_cpu", |
| 66 | ), |
| 67 | reason="CPU not available or DeepSeek-V4 CPU kernels not built in", |
| 68 | ) |
| 69 | |
| 70 | |
| 71 | def _apply_gptj_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor): |
| 72 | x_even = x[..., 0::2] |
| 73 | x_odd = x[..., 1::2] |
| 74 | out_even = x_even * cos - x_odd * sin |
| 75 | out_odd = x_even * sin + x_odd * cos |
| 76 | return torch.stack((out_even, out_odd), dim=-1).flatten(-2) |
| 77 | |
| 78 | |
| 79 | def _make_cache(num_blocks: int, block_size: int) -> torch.Tensor: |
| 80 | block_bytes = block_size * TOKEN_DATA_BYTES + block_size * SCALE_BYTES_PER_TOKEN |
| 81 | return torch.zeros(num_blocks, block_bytes, dtype=torch.uint8) |
| 82 | |
| 83 | |
| 84 | def _decode_token_regions( |
| 85 | cache: torch.Tensor, slot_mapping: torch.Tensor, block_size: int |
| 86 | ): |
| 87 | """Slice out, per valid token, the raw NoPE-fp8 / RoPE-bf16 / scale byte |
| 88 | regions this layout defines.""" |
| 89 | valid = slot_mapping >= 0 |
| 90 | idx = valid.nonzero(as_tuple=True)[0] |
| 91 | slots = slot_mapping[idx] |
| 92 | block_idx = torch.div(slots, block_size, rounding_mode="floor") |
| 93 | pos_in_block = slots % block_size |
| 94 | |
| 95 | nope, rope, scale = [], [], [] |
| 96 | for b, p in zip(block_idx.tolist(), pos_in_block.tolist()): |
| 97 | token_off = p * TOKEN_DATA_BYTES |
| 98 | nope.append(cache[b, token_off : token_off + NOPE_DIM]) |
| 99 | rope.append(cache[b, token_off + NOPE_DIM : token_off + TOKEN_DATA_BYTES]) |
| 100 | scale_off = block_size * TOKEN_DATA_BYTES + p * SCALE_BYTES_PER_TOKEN |
| 101 | scale.append(cache[b, scale_off : scale_off + SCALE_BYTES_PER_TOKEN - 1]) |
| 102 | |
| 103 | if not nope: |
| 104 | empty = lambda n: torch.empty(0, n, dtype=torch.uint8) |
| 105 | return empty(NOPE_DIM), empty(ROPE_DIM * 2), empty(SCALE_BYTES_PER_TOKEN - 1) |
| 106 | return torch.stack(nope), torch.stack(rope), torch.stack(scale) |
| 107 | |
| 108 | |
| 109 | def _assert_cache_parity(cache_fused, cache_ref, slot_mapping, block_size): |
| 110 | """NoPE fp8 + the UE8M0 scale byte are a deterministic quant of an |
| 111 | un-rotated bf16 value, so they match tightly; the RoPE region is rotated |
| 112 | in fp32 before its bf16 store and eager/kernel can round a tie to |
| 113 | opposite sides, so allow <=1 ULP there.""" |
| 114 | nope_ref, rope_ref, scale_ref = _decode_token_regions( |
| 115 | cache_ref, slot_mapping, block_size |
| 116 | ) |
| 117 | nope_fused, rope_fused, scale_fused = _decode_token_regions( |
| 118 | cache_fused, slot_mapping, block_size |
| 119 | ) |
| 120 | if nope_ref.numel() == 0: |
| 121 | return |
| 122 | |
| 123 | max_ulp = int(fp8_ulp_distance(nope_fused, nope_ref).max().item()) |
| 124 | assert max_ulp <= 1, f"NoPE fp8 differs by {max_ulp} ULP (>1)" |
| 125 | |
| 126 | rope_ref_bf16 = rope_ref.contiguous().view(torch.bfloat16) |
| 127 | rope_fused_bf16 = rope_fused.contiguous().view(torch.bfloat16) |
| 128 | max_ulp_rope = int(bf16_ulp_distance(rope_fused_bf16, rope_ref_bf16).max().item()) |
| 129 | assert max_ulp_rope <= 1, f"RoPE bf16 differs by {max_ulp_rope} ULP (>1)" |
| 130 | |
| 131 | scale_diff = (scale_fused.int() - scale_ref.int()).abs().max().item() |
| 132 | assert scale_diff <= 1, f"scale byte differs by {scale_diff} (>1)" |
| 133 | |
| 134 | |
| 135 | def _make_indexer_cache(num_blocks: int, block_size: int) -> torch.Tensor: |
| 136 | block_bytes = block_size * INDEXER_HEAD_DIM + block_size * INDEXER_SCALE_BYTES |
| 137 | return torch.zeros(num_blocks, block_bytes, dtype=torch.uint8) |
| 138 | |
| 139 | |
| 140 | def _assert_indexer_cache_parity(cache_fused, cache_ref, slot_mapping, block_size): |
| 141 | valid = slot_mapping >= 0 |
| 142 | idx = valid.nonzero(as_tuple=True)[0] |
| 143 | if idx.numel() == 0: |
| 144 | return |
| 145 | slots = slot_mapping[idx] |
| 146 | block_idx = torch.div(slots, block_size, rounding_mode="floor") |
| 147 | pos_in_block = slots % block_size |
| 148 | |
| 149 | fp8_ref, scale_ref = [], [] |
| 150 | fp8_fused, scale_fused = [], [] |
| 151 | for b, p in zip(block_idx.tolist(), pos_in_block.tolist()): |
| 152 | token_off = p * INDEXER_HEAD_DIM |
| 153 | scale_off = block_size * INDEXER_HEAD_DIM + p * INDEXER_SCALE_BYTES |
| 154 | fp8_ref.append(cache_ref[b, token_off : token_off + INDEXER_HEAD_DIM]) |
| 155 | fp8_fused.append(cache_fused[b, token_off : token_off + INDEXER_HEAD_DIM]) |
| 156 | scale_ref.append(cache_ref[b, scale_off : scale_off + INDEXER_SCALE_BYTES]) |
| 157 | scale_fused.append(cache_fused[b, scale_off : scale_off + INDEXER_SCALE_BYTES]) |
| 158 | |
| 159 | fp8_ref_t = torch.stack(fp8_ref).view(torch.float8_e4m3fn) |
| 160 | fp8_fused_t = torch.stack(fp8_fused).view(torch.float8_e4m3fn) |
| 161 | torch.testing.assert_close(fp8_fused_t.float(), fp8_ref_t.float()) |
| 162 | |
| 163 | scale_ref_t = torch.stack(scale_ref).view(torch.float32) |
| 164 | scale_fused_t = torch.stack(scale_fused).view(torch.float32) |
| 165 | torch.testing.assert_close(scale_fused_t, scale_ref_t) |
| 166 | |
| 167 | |
| 168 | # --------------------------------------------------------------------------- |
| 169 | # save_partial_states_cpu / compress_norm_rope_store_cpu (compressor.cpp) |
| 170 | # --------------------------------------------------------------------------- |
| 171 | # |
| 172 | # The compressor's state cache is pooled with the main KV cache's physical |
| 173 | # pages and page-aligned, so it is genuinely 3D with a block-to-block stride |
| 174 | # that need not equal block_size * width -- built here via torch.as_strided |
| 175 | # over an over-sized flat buffer rather than a plain contiguous tensor, so a |
| 176 | # regression back to flattening via .view() would be caught. |
| 177 | |
| 178 | |
| 179 | def _make_padded_block_cache( |
| 180 | num_blocks: int, block_size: int, width: int, pad_elems: int |
| 181 | ) -> torch.Tensor: |
| 182 | page_stride = block_size * width + pad_elems |
| 183 | flat = torch.randn(num_blocks, page_stride, dtype=torch.float32) |
| 184 | return torch.as_strided( |
| 185 | flat, (num_blocks, block_size, width), (page_stride, width, 1) |
| 186 | ) |
| 187 | |
| 188 | |
| 189 | def _dup_padded_block_cache( |
| 190 | num_blocks: int, block_size: int, width: int, pad_elems: int |
| 191 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 192 | """Two independently-stored padded caches with identical values, for ops |
| 193 | that mutate the cache in place.""" |
| 194 | page_stride = block_size * width + pad_elems |
| 195 | template = torch.randn(num_blocks, page_stride, dtype=torch.float32) |
| 196 | |
| 197 | def _view(flat: torch.Tensor) -> torch.Tensor: |
| 198 | return torch.as_strided( |
| 199 | flat, (num_blocks, block_size, width), (page_stride, width, 1) |
| 200 | ) |
| 201 | |
| 202 | return _view(template.clone()), _view(template.clone()) |
| 203 | |
| 204 | |
| 205 | def _run_save_partial_states_eager( |
| 206 | state_cache, kv, score, ape, positions, slot_mapping, compress_ratio |
| 207 | ): |
| 208 | """Reference: state_cache[block_idx, pos_in_block, :W] = kv; |
| 209 | state_cache[..., W:] = score + ape[position % compress_ratio]; skip |
| 210 | slot < 0. Mutates state_cache in place, matching the op's contract.""" |
| 211 | state_width = kv.shape[-1] |
| 212 | block_size = state_cache.shape[1] |
| 213 | valid = slot_mapping >= 0 |
| 214 | idx = valid.nonzero(as_tuple=True)[0] |
| 215 | slots = slot_mapping[idx] |
| 216 | block_idx = slots // block_size |
| 217 | pos_in_block = slots % block_size |
| 218 | |
| 219 | state_cache[block_idx, pos_in_block, :state_width] = kv[idx] |
| 220 | ape_row = ape[positions[idx] % compress_ratio] |
| 221 | state_cache[block_idx, pos_in_block, state_width:] = score[idx] + ape_row |
| 222 | |
| 223 | |
| 224 | @pytest.mark.parametrize("pad_elems", [0, 24]) |
| 225 | @pytest.mark.parametrize("num_tokens", [1, 17]) |
| 226 | def test_save_partial_states_cpu_matches_eager(num_tokens: int, pad_elems: int): |
| 227 | torch.manual_seed(0) |
| 228 | compress_ratio = 4 |
| 229 | block_size = 4 |
| 230 | state_width = 512 |
| 231 | width = 2 * state_width |
| 232 | num_blocks = (num_tokens + 4 + block_size - 1) // block_size + 1 |
| 233 | num_slots = num_blocks * block_size |
| 234 | |
| 235 | kv = torch.randn(num_tokens, state_width, dtype=torch.float32) |
| 236 | score = torch.randn(num_tokens, state_width, dtype=torch.float32) |
| 237 | ape = torch.randn(compress_ratio, state_width, dtype=torch.float32) |
| 238 | positions = torch.randint(0, 4096, (num_tokens,), dtype=torch.int64) |
| 239 | |
| 240 | ref_cache, fused_cache = _dup_padded_block_cache( |
| 241 | num_blocks, block_size, width, pad_elems |
| 242 | ) |
| 243 | |
| 244 | slot_mapping = torch.randperm(num_slots)[:num_tokens].to(torch.int64) |
| 245 | if num_tokens > 1: |
| 246 | slot_mapping[-1] = -1 # exercise the "skip invalid slot" branch |
| 247 | |
| 248 | _run_save_partial_states_eager( |
| 249 | ref_cache, kv, score, ape, positions, slot_mapping, compress_ratio |
| 250 | ) |
| 251 | save_partial_states_cpu(kv, score, ape, positions, fused_cache, slot_mapping) |
| 252 | |
| 253 | torch.testing.assert_close(fused_cache, ref_cache) |
| 254 | |
| 255 | |
| 256 | def _run_compress_eager( |
| 257 | state_cache, |
| 258 | gather_slots, |
| 259 | positions, |
| 260 | kv_slot_mapping, |
| 261 | rms_norm_weight, |
| 262 | rms_norm_eps, |
| 263 | cos_sin_cache, |
| 264 | kv_cache, |
| 265 | kv_cache_block_size, |
| 266 | compress_ratio, |
| 267 | ): |
| 268 | """Reference for the per-channel-softmax-pooling + RMSNorm + NoPE-UE8M0 |
| 269 | quant + GPT-J RoPE math writing the main (head_dim=512) fp8_ds_mla |
| 270 | layout.""" |
| 271 | num_tokens, window = gather_slots.shape |
| 272 | state_width = state_cache.shape[-1] // 2 |
| 273 | state_block_size = state_cache.shape[1] |
| 274 | num_quant_blocks = NOPE_DIM // QUANT_BLOCK |
| 275 | |
| 276 | w_idx = torch.arange(window) |
| 277 | head_offset = torch.where(w_idx >= compress_ratio, HEAD_DIM, 0) |
| 278 | |
| 279 | for t in range(num_tokens): |
| 280 | position = int(positions[t]) |
| 281 | if (position + 1) % compress_ratio != 0: |
| 282 | continue |
| 283 | kv_slot = int(kv_slot_mapping[t]) |
| 284 | if kv_slot < 0: |
| 285 | continue |
| 286 | |
| 287 | slots = gather_slots[t] |
| 288 | valid_w = slots >= 0 |
| 289 | safe_slots = slots.clamp(min=0) |
| 290 | block_idx = safe_slots // state_block_size |
| 291 | pos_in_block = safe_slots % state_block_size |
| 292 | |
| 293 | rows = state_cache[block_idx, pos_in_block] # [window, 2*state_width] |
| 294 | col = head_offset.unsqueeze(-1) + torch.arange(HEAD_DIM) |
| 295 | kv_vals = torch.gather(rows, 1, col) |
| 296 | score_col = state_width + col |
| 297 | score_vals = torch.gather(rows, 1, score_col) |
| 298 | score_vals = torch.where( |
| 299 | valid_w.unsqueeze(-1), score_vals, torch.full_like(score_vals, -torch.inf) |
| 300 | ) |
| 301 | |
| 302 | weights = torch.softmax(score_vals, dim=0) |
| 303 | compressed = (kv_vals * weights).sum(dim=0) # [HEAD_DIM] |
| 304 | |
| 305 | variance = compressed.pow(2).mean() |
| 306 | normed = compressed * torch.rsqrt(variance + rms_norm_eps) * rms_norm_weight |
| 307 | |
| 308 | quant_input = normed.to(torch.bfloat16).to(torch.float32) |
| 309 | |
| 310 | compressed_pos = (position // compress_ratio) * compress_ratio |
| 311 | cos = cos_sin_cache[compressed_pos, : ROPE_DIM // 2] |
| 312 | sin = cos_sin_cache[compressed_pos, ROPE_DIM // 2 :] |
| 313 | rotated_rope = _apply_gptj_rope(normed[NOPE_DIM:].clone(), cos, sin) |
| 314 | |
| 315 | nope_vals = quant_input[:NOPE_DIM].view(num_quant_blocks, QUANT_BLOCK) |
| 316 | absmax = nope_vals.abs().amax(dim=-1).clamp(min=1e-4) |
| 317 | exponent = torch.ceil(torch.log2(absmax / FP8_MAX)) |
| 318 | inv_scale = torch.exp2(-exponent) |
| 319 | scaled = (nope_vals * inv_scale.unsqueeze(-1)).clamp(-FP8_MAX, FP8_MAX) |
| 320 | nope_fp8_bytes = scaled.to(torch.float8_e4m3fn).reshape(-1).view(torch.uint8) |
| 321 | scale_bytes = (exponent + 127.0).clamp(0, 255).to(torch.uint8) |
| 322 | rope_bf16_bytes = rotated_rope.to(torch.bfloat16).contiguous().view(torch.uint8) |
| 323 | |
| 324 | block_idx_kv = kv_slot // kv_cache_block_size |
| 325 | pos_in_block_kv = kv_slot % kv_cache_block_size |
| 326 | token_off = pos_in_block_kv * TOKEN_DATA_BYTES |
| 327 | kv_cache[block_idx_kv, token_off : token_off + NOPE_DIM] = nope_fp8_bytes |
| 328 | kv_cache[block_idx_kv, token_off + NOPE_DIM : token_off + TOKEN_DATA_BYTES] = ( |
| 329 | rope_bf16_bytes |
| 330 | ) |
| 331 | scale_off = ( |
| 332 | kv_cache_block_size * TOKEN_DATA_BYTES |
| 333 | + pos_in_block_kv * SCALE_BYTES_PER_TOKEN |
| 334 | ) |
| 335 | kv_cache[block_idx_kv, scale_off : scale_off + num_quant_blocks] = scale_bytes |
| 336 | kv_cache[block_idx_kv, scale_off + num_quant_blocks] = 0 |
| 337 | |
| 338 | |
| 339 | def _make_compress_inputs( |
| 340 | num_tokens: int, |
| 341 | compress_ratio: int, |
| 342 | overlap: bool, |
| 343 | pad_elems: int = 24, |
| 344 | max_pos: int = 4096, |
| 345 | invalid_window_frac: float = 0.1, |
| 346 | head_dim: int = HEAD_DIM, |
| 347 | ): |
| 348 | torch.manual_seed(0) |
| 349 | coff = 1 + int(overlap) |
| 350 | state_width = coff * head_dim |
| 351 | window = coff * compress_ratio |
| 352 | state_block_size = 4 if compress_ratio == 4 else 8 |
| 353 | num_state_slots = num_tokens * window + 8 |
| 354 | num_state_blocks = (num_state_slots + state_block_size - 1) // state_block_size + 1 |
| 355 | |
| 356 | state_cache = _make_padded_block_cache( |
| 357 | num_state_blocks, state_block_size, 2 * state_width, pad_elems |
| 358 | ) |
| 359 | gather_slots = torch.randint( |
| 360 | 0, num_state_blocks * state_block_size, (num_tokens, window) |
| 361 | ).to(torch.int64) |
| 362 | invalid_mask = torch.rand(num_tokens, window) < invalid_window_frac |
| 363 | # Never invalidate every column of a row (would force max=-inf -> NaN). |
| 364 | invalid_mask[:, -1] = False |
| 365 | gather_slots = torch.where( |
| 366 | invalid_mask, torch.full_like(gather_slots, -1), gather_slots |
| 367 | ) |
| 368 | |
| 369 | base = (torch.arange(num_tokens) + 1) * compress_ratio |
| 370 | is_boundary = torch.arange(num_tokens) % 2 == 0 |
| 371 | positions = torch.where(is_boundary, base - 1, base + 1).to(torch.int64) |
| 372 | |
| 373 | kv_slot_mapping = torch.arange(num_tokens, dtype=torch.int64) |
| 374 | if num_tokens > 1: |
| 375 | kv_slot_mapping[1] = -1 # exercise the "skip invalid output slot" branch |
| 376 | |
| 377 | rms_norm_weight = torch.randn(head_dim, dtype=torch.float32) |
| 378 | cos_sin_cache = make_cos_sin_cache(max_pos, ROPE_DIM, torch.float32, "cpu") |
| 379 | |
| 380 | return ( |
| 381 | state_cache, |
| 382 | gather_slots, |
| 383 | positions, |
| 384 | kv_slot_mapping, |
| 385 | rms_norm_weight, |
| 386 | cos_sin_cache, |
| 387 | is_boundary, |
| 388 | ) |
| 389 | |
| 390 | |
| 391 | @pytest.mark.parametrize("compress_ratio,overlap", [(128, False), (4, True)]) |
| 392 | @pytest.mark.parametrize("num_tokens", [1, 17]) |
| 393 | def test_compress_norm_rope_store_cpu_matches_eager( |
| 394 | compress_ratio: int, overlap: bool, num_tokens: int |
| 395 | ): |
| 396 | eps = 1e-6 |
| 397 | block_size = 16 |
| 398 | ( |
| 399 | state_cache, |
| 400 | gather_slots, |
| 401 | positions, |
| 402 | kv_slot_mapping, |
| 403 | rms_norm_weight, |
| 404 | cos_sin_cache, |
| 405 | is_boundary, |
| 406 | ) = _make_compress_inputs(num_tokens, compress_ratio, overlap) |
| 407 | |
| 408 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 409 | kv_cache_ref = _make_cache(num_blocks, block_size) |
| 410 | kv_cache_fused = kv_cache_ref.clone() |
| 411 | |
| 412 | _run_compress_eager( |
| 413 | state_cache, |
| 414 | gather_slots, |
| 415 | positions, |
| 416 | kv_slot_mapping, |
| 417 | rms_norm_weight, |
| 418 | eps, |
| 419 | cos_sin_cache, |
| 420 | kv_cache_ref, |
| 421 | block_size, |
| 422 | compress_ratio, |
| 423 | ) |
| 424 | compress_norm_rope_store_cpu( |
| 425 | state_cache, |
| 426 | gather_slots, |
| 427 | positions, |
| 428 | kv_slot_mapping, |
| 429 | rms_norm_weight, |
| 430 | eps, |
| 431 | cos_sin_cache, |
| 432 | kv_cache_fused, |
| 433 | block_size, |
| 434 | compress_ratio, |
| 435 | ) |
| 436 | |
| 437 | # Only tokens that hit the compression boundary (and have a valid output |
| 438 | # slot) were written -- restrict parity checks to those. |
| 439 | written_slot_mapping = torch.where( |
| 440 | is_boundary, kv_slot_mapping, torch.full_like(kv_slot_mapping, -1) |
| 441 | ) |
| 442 | _assert_cache_parity(kv_cache_fused, kv_cache_ref, written_slot_mapping, block_size) |
| 443 | |
| 444 | |
| 445 | def test_compress_norm_rope_store_cpu_non_contiguous_kv_cache_row_stride(): |
| 446 | """The main (output) kv_cache is a separate physical buffer from the |
| 447 | state cache and can independently be a strided view.""" |
| 448 | eps = 1e-6 |
| 449 | compress_ratio, overlap = 128, False |
| 450 | num_tokens = 5 |
| 451 | block_size = 16 |
| 452 | cache_pad = 32 |
| 453 | |
| 454 | ( |
| 455 | state_cache, |
| 456 | gather_slots, |
| 457 | positions, |
| 458 | kv_slot_mapping, |
| 459 | rms_norm_weight, |
| 460 | cos_sin_cache, |
| 461 | is_boundary, |
| 462 | ) = _make_compress_inputs(num_tokens, compress_ratio, overlap) |
| 463 | |
| 464 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 465 | kv_cache_ref = _make_cache(num_blocks, block_size) |
| 466 | block_bytes = kv_cache_ref.shape[1] |
| 467 | padded_cache_ref = torch.zeros( |
| 468 | num_blocks, block_bytes + cache_pad, dtype=torch.uint8 |
| 469 | ) |
| 470 | kv_cache_ref_view = padded_cache_ref[:, :block_bytes] |
| 471 | kv_cache_ref_view.copy_(kv_cache_ref) |
| 472 | padded_cache_fused = padded_cache_ref.clone() |
| 473 | kv_cache_fused_view = padded_cache_fused[:, :block_bytes] |
| 474 | |
| 475 | _run_compress_eager( |
| 476 | state_cache, |
| 477 | gather_slots, |
| 478 | positions, |
| 479 | kv_slot_mapping, |
| 480 | rms_norm_weight, |
| 481 | eps, |
| 482 | cos_sin_cache, |
| 483 | kv_cache_ref_view, |
| 484 | block_size, |
| 485 | compress_ratio, |
| 486 | ) |
| 487 | compress_norm_rope_store_cpu( |
| 488 | state_cache, |
| 489 | gather_slots, |
| 490 | positions, |
| 491 | kv_slot_mapping, |
| 492 | rms_norm_weight, |
| 493 | eps, |
| 494 | cos_sin_cache, |
| 495 | kv_cache_fused_view, |
| 496 | block_size, |
| 497 | compress_ratio, |
| 498 | ) |
| 499 | |
| 500 | written_slot_mapping = torch.where( |
| 501 | is_boundary, kv_slot_mapping, torch.full_like(kv_slot_mapping, -1) |
| 502 | ) |
| 503 | _assert_cache_parity( |
| 504 | kv_cache_fused_view, kv_cache_ref_view, written_slot_mapping, block_size |
| 505 | ) |
| 506 | |
| 507 | |
| 508 | def _run_compress_indexer_eager( |
| 509 | state_cache, |
| 510 | gather_slots, |
| 511 | positions, |
| 512 | kv_slot_mapping, |
| 513 | rms_norm_weight, |
| 514 | rms_norm_eps, |
| 515 | cos_sin_cache, |
| 516 | kv_cache, |
| 517 | kv_cache_block_size, |
| 518 | compress_ratio, |
| 519 | ): |
| 520 | """Same softmax-pool + RMSNorm + GPT-J RoPE prologue as the head=512 |
| 521 | path, but the indexer's cache quantizes all 128 post-RoPE dims as one |
| 522 | FP8 block with a single raw fp32 scale per token (matching |
| 523 | _indexer_k_quant_and_cache_cpu's layout), not 7 separate UE8M0 blocks.""" |
| 524 | num_tokens, window = gather_slots.shape |
| 525 | state_width = state_cache.shape[-1] // 2 |
| 526 | state_block_size = state_cache.shape[1] |
| 527 | |
| 528 | w_idx = torch.arange(window) |
| 529 | head_offset = torch.where(w_idx >= compress_ratio, INDEXER_HEAD_DIM, 0) |
| 530 | |
| 531 | for t in range(num_tokens): |
| 532 | position = int(positions[t]) |
| 533 | if (position + 1) % compress_ratio != 0: |
| 534 | continue |
| 535 | kv_slot = int(kv_slot_mapping[t]) |
| 536 | if kv_slot < 0: |
| 537 | continue |
| 538 | |
| 539 | slots = gather_slots[t] |
| 540 | valid_w = slots >= 0 |
| 541 | safe_slots = slots.clamp(min=0) |
| 542 | block_idx = safe_slots // state_block_size |
| 543 | pos_in_block = safe_slots % state_block_size |
| 544 | |
| 545 | rows = state_cache[block_idx, pos_in_block] # [window, 2*state_width] |
| 546 | col = head_offset.unsqueeze(-1) + torch.arange(INDEXER_HEAD_DIM) |
| 547 | kv_vals = torch.gather(rows, 1, col) |
| 548 | score_col = state_width + col |
| 549 | score_vals = torch.gather(rows, 1, score_col) |
| 550 | score_vals = torch.where( |
| 551 | valid_w.unsqueeze(-1), score_vals, torch.full_like(score_vals, -torch.inf) |
| 552 | ) |
| 553 | |
| 554 | weights = torch.softmax(score_vals, dim=0) |
| 555 | compressed = (kv_vals * weights).sum(dim=0) # [INDEXER_HEAD_DIM] |
| 556 | |
| 557 | variance = compressed.pow(2).mean() |
| 558 | normed = compressed * torch.rsqrt(variance + rms_norm_eps) * rms_norm_weight |
| 559 | |
| 560 | compressed_pos = (position // compress_ratio) * compress_ratio |
| 561 | cos = cos_sin_cache[compressed_pos, : ROPE_DIM // 2] |
| 562 | sin = cos_sin_cache[compressed_pos, ROPE_DIM // 2 :] |
| 563 | rotated_rope = _apply_gptj_rope(normed[INDEXER_NOPE_DIM:].clone(), cos, sin) |
| 564 | full = torch.cat((normed[:INDEXER_NOPE_DIM], rotated_rope)) |
| 565 | |
| 566 | quant_input = full.to(torch.bfloat16).to(torch.float32) |
| 567 | absmax = quant_input.abs().amax().clamp(min=1e-4) |
| 568 | exponent = torch.ceil(torch.log2(absmax / FP8_MAX)) |
| 569 | scale = torch.exp2(exponent) |
| 570 | scaled = (quant_input / scale).clamp(-FP8_MAX, FP8_MAX) |
| 571 | fp8_bytes = scaled.to(torch.float8_e4m3fn).view(torch.uint8) |
| 572 | |
| 573 | block_idx_kv = kv_slot // kv_cache_block_size |
| 574 | pos_in_block_kv = kv_slot % kv_cache_block_size |
| 575 | token_off = pos_in_block_kv * INDEXER_HEAD_DIM |
| 576 | kv_cache[block_idx_kv, token_off : token_off + INDEXER_HEAD_DIM] = fp8_bytes |
| 577 | scale_off = ( |
| 578 | kv_cache_block_size * INDEXER_HEAD_DIM |
| 579 | + pos_in_block_kv * INDEXER_SCALE_BYTES |
| 580 | ) |
| 581 | kv_cache[block_idx_kv, scale_off : scale_off + INDEXER_SCALE_BYTES] = ( |
| 582 | scale.reshape(1).view(torch.uint8) |
| 583 | ) |
| 584 | |
| 585 | |
| 586 | @pytest.mark.parametrize("compress_ratio,overlap", [(128, False), (4, True)]) |
| 587 | @pytest.mark.parametrize("num_tokens", [1, 17]) |
| 588 | def test_compress_norm_rope_store_indexer_cpu_matches_eager( |
| 589 | compress_ratio: int, overlap: bool, num_tokens: int |
| 590 | ): |
| 591 | eps = 1e-6 |
| 592 | block_size = 16 |
| 593 | ( |
| 594 | state_cache, |
| 595 | gather_slots, |
| 596 | positions, |
| 597 | kv_slot_mapping, |
| 598 | rms_norm_weight, |
| 599 | cos_sin_cache, |
| 600 | is_boundary, |
| 601 | ) = _make_compress_inputs( |
| 602 | num_tokens, compress_ratio, overlap, head_dim=INDEXER_HEAD_DIM |
| 603 | ) |
| 604 | |
| 605 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 606 | kv_cache_ref = _make_indexer_cache(num_blocks, block_size) |
| 607 | kv_cache_fused = kv_cache_ref.clone() |
| 608 | |
| 609 | _run_compress_indexer_eager( |
| 610 | state_cache, |
| 611 | gather_slots, |
| 612 | positions, |
| 613 | kv_slot_mapping, |
| 614 | rms_norm_weight, |
| 615 | eps, |
| 616 | cos_sin_cache, |
| 617 | kv_cache_ref, |
| 618 | block_size, |
| 619 | compress_ratio, |
| 620 | ) |
| 621 | compress_norm_rope_store_indexer_cpu( |
| 622 | state_cache, |
| 623 | gather_slots, |
| 624 | positions, |
| 625 | kv_slot_mapping, |
| 626 | rms_norm_weight, |
| 627 | eps, |
| 628 | cos_sin_cache, |
| 629 | kv_cache_fused, |
| 630 | block_size, |
| 631 | compress_ratio, |
| 632 | ) |
| 633 | |
| 634 | written_slot_mapping = torch.where( |
| 635 | is_boundary, kv_slot_mapping, torch.full_like(kv_slot_mapping, -1) |
| 636 | ) |
| 637 | _assert_indexer_cache_parity( |
| 638 | kv_cache_fused, kv_cache_ref, written_slot_mapping, block_size |
| 639 | ) |
| 640 | |
| 641 | |
| 642 | # --------------------------------------------------------------------------- |
| 643 | # flash_mla_with_kvcache_cpu (flash_mla.cpp): fused sparse MQA decode attn |
| 644 | # --------------------------------------------------------------------------- |
| 645 | |
| 646 | |
| 647 | def _pack_flash_mla_cache(kv_f32: torch.Tensor, block_size: int): |
| 648 | """Quantize kv_f32 ([N, HEAD_DIM]) into the fp8_ds_mla byte layout and |
| 649 | return (cache, dequant_ref) -- the exact post-quantization value the |
| 650 | kernel's dequant should reproduce. Slot id == row index, packed |
| 651 | contiguously from block 0.""" |
| 652 | n = kv_f32.shape[0] |
| 653 | num_blocks = max(1, -(-n // block_size)) |
| 654 | cache = _make_cache(num_blocks, block_size) |
| 655 | if n == 0: |
| 656 | return cache, kv_f32.new_zeros(0, HEAD_DIM) |
| 657 | |
| 658 | num_quant_blocks = NOPE_DIM // QUANT_BLOCK |
| 659 | nope = kv_f32[:, :NOPE_DIM].reshape(n, num_quant_blocks, QUANT_BLOCK) |
| 660 | absmax = nope.abs().amax(dim=-1).clamp(min=1e-4) |
| 661 | exponent = torch.ceil(torch.log2(absmax / FP8_MAX)) |
| 662 | inv_scale = torch.exp2(-exponent) |
| 663 | scaled = (nope * inv_scale.unsqueeze(-1)).clamp(-FP8_MAX, FP8_MAX) |
| 664 | nope_fp8 = scaled.to(torch.float8_e4m3fn) |
| 665 | scale = torch.exp2(exponent) |
| 666 | nope_dequant = (nope_fp8.float() * scale.unsqueeze(-1)).reshape(n, NOPE_DIM) |
| 667 | |
| 668 | rope_bf16 = kv_f32[:, NOPE_DIM:].to(torch.bfloat16) |
| 669 | dequant_ref = torch.cat([nope_dequant, rope_bf16.float()], dim=-1) |
| 670 | |
| 671 | slot_ids = torch.arange(n) |
| 672 | block_idx = slot_ids // block_size |
| 673 | pos_in_block = slot_ids % block_size |
| 674 | |
| 675 | nope_fp8_bytes = nope_fp8.reshape(n, NOPE_DIM).contiguous().view(torch.uint8) |
| 676 | rope_bf16_bytes = rope_bf16.contiguous().view(torch.uint8) |
| 677 | scale_bytes = (exponent + 127.0).clamp(0, 255).to(torch.uint8) |
| 678 | |
| 679 | arange_nope = torch.arange(NOPE_DIM) |
| 680 | arange_rope = torch.arange(ROPE_DIM * 2) |
| 681 | arange_scale = torch.arange(num_quant_blocks) |
| 682 | token_base = pos_in_block.unsqueeze(-1) * TOKEN_DATA_BYTES |
| 683 | row_idx = block_idx.unsqueeze(-1) |
| 684 | |
| 685 | cache[row_idx, token_base + arange_nope] = nope_fp8_bytes |
| 686 | cache[row_idx, token_base + NOPE_DIM + arange_rope] = rope_bf16_bytes |
| 687 | scale_base = ( |
| 688 | block_size * TOKEN_DATA_BYTES |
| 689 | + pos_in_block.unsqueeze(-1) * SCALE_BYTES_PER_TOKEN |
| 690 | ) |
| 691 | cache[row_idx, scale_base + arange_scale] = scale_bytes |
| 692 | cache[block_idx, scale_base.squeeze(-1) + num_quant_blocks] = 0 |
| 693 | |
| 694 | return cache, dequant_ref |
| 695 | |
| 696 | |
| 697 | def _flash_mla_eager( |
| 698 | q: torch.Tensor, |
| 699 | window_dequant: torch.Tensor, |
| 700 | window_slots: torch.Tensor, |
| 701 | compressed_dequant: torch.Tensor, |
| 702 | compressed_slots: torch.Tensor, |
| 703 | attn_sink: torch.Tensor, |
| 704 | scale: float, |
| 705 | ) -> torch.Tensor: |
| 706 | """MQA attention: gather the (already-dequantized) window+compressed KV |
| 707 | rows referenced by *_slots (-1 = invalid/masked), append a learned |
| 708 | per-head sink logit with zero value contribution, softmax, weighted |
| 709 | sum.""" |
| 710 | num_tokens, num_heads, _ = q.shape |
| 711 | q_f32 = q.float() |
| 712 | |
| 713 | def gather(dequant, slots): |
| 714 | valid = slots >= 0 |
| 715 | safe = slots.clamp(min=0) |
| 716 | kv = dequant[safe] # [T, K, D] |
| 717 | return kv, valid |
| 718 | |
| 719 | win_kv, win_valid = gather(window_dequant, window_slots) |
| 720 | if compressed_slots.shape[1] > 0: |
| 721 | comp_kv, comp_valid = gather(compressed_dequant, compressed_slots) |
| 722 | kv = torch.cat([win_kv, comp_kv], dim=1) |
| 723 | valid = torch.cat([win_valid, comp_valid], dim=1) |
| 724 | else: |
| 725 | kv = win_kv |
| 726 | valid = win_valid |
| 727 | |
| 728 | scores = torch.einsum("thd,tkd->thk", q_f32, kv) * scale |
| 729 | scores = scores.masked_fill(~valid.unsqueeze(1), float("-inf")) |
| 730 | sink = attn_sink.view(1, num_heads, 1).expand(num_tokens, num_heads, 1) |
| 731 | scores_full = torch.cat([scores, sink], dim=-1) |
| 732 | probs = torch.softmax(scores_full, dim=-1) |
| 733 | out = torch.einsum("thk,tkd->thd", probs[..., :-1], kv) |
| 734 | return out.to(q.dtype) |
| 735 | |
| 736 | |
| 737 | @pytest.mark.parametrize("num_heads", [1, 64]) |
| 738 | @pytest.mark.parametrize("num_tokens", [1, 17]) |
| 739 | @pytest.mark.parametrize("swa_only", [True, False]) |
| 740 | def test_flash_mla_cpu_kernel_matches_eager( |
| 741 | num_heads: int, num_tokens: int, swa_only: bool |
| 742 | ): |
| 743 | torch.manual_seed(0) |
| 744 | dtype = torch.bfloat16 |
| 745 | window_block_size = 16 |
| 746 | compressed_block_size = 8 |
| 747 | num_window = 6 |
| 748 | num_compressed = 0 if swa_only else 5 |
| 749 | |
| 750 | q = torch.randn(num_tokens, num_heads, HEAD_DIM, dtype=dtype) |
| 751 | scale = 1.0 / (HEAD_DIM**0.5) |
| 752 | attn_sink = torch.randn(num_heads, dtype=torch.float32) |
| 753 | |
| 754 | num_window_slots = num_tokens * num_window + window_block_size |
| 755 | window_kv_f32 = torch.randn(num_window_slots, HEAD_DIM) |
| 756 | window_cache, window_dequant = _pack_flash_mla_cache( |
| 757 | window_kv_f32, window_block_size |
| 758 | ) |
| 759 | window_slots = torch.randint(0, num_window_slots, (num_tokens, num_window)) |
| 760 | window_slots[:, 0] = -1 # exercise invalid-slot masking |
| 761 | |
| 762 | if num_compressed > 0: |
| 763 | num_compressed_slots = num_tokens * num_compressed + compressed_block_size |
| 764 | compressed_kv_f32 = torch.randn(num_compressed_slots, HEAD_DIM) |
| 765 | compressed_cache, compressed_dequant = _pack_flash_mla_cache( |
| 766 | compressed_kv_f32, compressed_block_size |
| 767 | ) |
| 768 | compressed_slots = torch.randint( |
| 769 | 0, num_compressed_slots, (num_tokens, num_compressed) |
| 770 | ) |
| 771 | if num_compressed > 1: |
| 772 | compressed_slots[:, -1] = -1 |
| 773 | else: |
| 774 | compressed_cache = window_cache |
| 775 | compressed_dequant = window_dequant.new_zeros(0, HEAD_DIM) |
| 776 | compressed_slots = window_slots.new_full((num_tokens, 0), -1) |
| 777 | |
| 778 | out_ref = _flash_mla_eager( |
| 779 | q, |
| 780 | window_dequant, |
| 781 | window_slots, |
| 782 | compressed_dequant, |
| 783 | compressed_slots, |
| 784 | attn_sink, |
| 785 | scale, |
| 786 | ) |
| 787 | out_kernel = q.new_empty(num_tokens, num_heads, HEAD_DIM) |
| 788 | flash_mla_with_kvcache_cpu( |
| 789 | out_kernel, |
| 790 | q, |
| 791 | window_cache, |
| 792 | window_slots, |
| 793 | window_block_size, |
| 794 | compressed_cache, |
| 795 | compressed_slots, |
| 796 | compressed_block_size, |
| 797 | attn_sink, |
| 798 | scale, |
| 799 | ) |
| 800 | |
| 801 | torch.testing.assert_close( |
| 802 | out_kernel.float(), out_ref.float(), rtol=2e-2, atol=2e-2 |
| 803 | ) |
| 804 | |
| 805 | |
| 806 | def test_flash_mla_cpu_kernel_all_invalid_falls_back_to_sink(): |
| 807 | """When every KV slot is masked out, softmax degenerates to the sink |
| 808 | logit alone and the output must be all-zero (no valid value rows).""" |
| 809 | torch.manual_seed(2) |
| 810 | dtype = torch.bfloat16 |
| 811 | num_tokens, num_heads, num_window = 3, 4, 5 |
| 812 | window_block_size = 16 |
| 813 | |
| 814 | q = torch.randn(num_tokens, num_heads, HEAD_DIM, dtype=dtype) |
| 815 | scale = 1.0 / (HEAD_DIM**0.5) |
| 816 | attn_sink = torch.randn(num_heads, dtype=torch.float32) |
| 817 | |
| 818 | window_kv_f32 = torch.randn(window_block_size, HEAD_DIM) |
| 819 | window_cache, _ = _pack_flash_mla_cache(window_kv_f32, window_block_size) |
| 820 | window_slots = torch.full((num_tokens, num_window), -1, dtype=torch.int64) |
| 821 | compressed_slots = window_slots.new_full((num_tokens, 0), -1) |
| 822 | |
| 823 | out = q.new_empty(num_tokens, num_heads, HEAD_DIM) |
| 824 | flash_mla_with_kvcache_cpu( |
| 825 | out, |
| 826 | q, |
| 827 | window_cache, |
| 828 | window_slots, |
| 829 | window_block_size, |
| 830 | window_cache, |
| 831 | compressed_slots, |
| 832 | 1, |
| 833 | attn_sink, |
| 834 | scale, |
| 835 | ) |
| 836 | assert out.abs().max().item() == 0.0 |
| 837 | |
| 838 | |
| 839 | # --------------------------------------------------------------------------- |
| 840 | # fused_qnorm_rope_kv_insert_cpu (store_cache.cpp) |
| 841 | # --------------------------------------------------------------------------- |
| 842 | |
| 843 | |
| 844 | def _qnorm_kv_insert_eager( |
| 845 | q, kv, positions, cache, slot_mapping, cos_sin_cache, q_head_padded, eps, bs |
| 846 | ): |
| 847 | """Reference: per-head weight-free RMSNorm + GPT-J RoPE on Q; GPT-J RoPE |
| 848 | + UE8M0 FP8 quant + paged insert on KV.""" |
| 849 | num_quant_blocks = NOPE_DIM // QUANT_BLOCK |
| 850 | |
| 851 | num_tokens_full, num_heads_q, _ = q.shape |
| 852 | out_dtype = q.dtype |
| 853 | |
| 854 | cos = cos_sin_cache[positions, : ROPE_DIM // 2] |
| 855 | sin = cos_sin_cache[positions, ROPE_DIM // 2 :] |
| 856 | |
| 857 | q_f32 = q.float() |
| 858 | var = q_f32.square().mean(dim=-1, keepdim=True) |
| 859 | q_f32 = q_f32 * torch.rsqrt(var + eps) |
| 860 | q_f32[..., NOPE_DIM:] = _apply_gptj_rope( |
| 861 | q_f32[..., NOPE_DIM:], cos.unsqueeze(1), sin.unsqueeze(1) |
| 862 | ) |
| 863 | q_out = q.new_zeros(num_tokens_full, q_head_padded, HEAD_DIM) |
| 864 | q_out[:, :num_heads_q, :] = q_f32.to(out_dtype) |
| 865 | |
| 866 | num_tokens_insert = slot_mapping.shape[0] |
| 867 | kv_f32 = kv[:num_tokens_insert].float() |
| 868 | kv_f32[..., NOPE_DIM:] = _apply_gptj_rope( |
| 869 | kv_f32[..., NOPE_DIM:], cos[:num_tokens_insert], sin[:num_tokens_insert] |
| 870 | ) |
| 871 | kv_bf16 = kv_f32.to(torch.bfloat16) |
| 872 | kv_f32 = kv_bf16.float() |
| 873 | |
| 874 | valid = slot_mapping >= 0 |
| 875 | if bool(valid.any()): |
| 876 | idx = valid.nonzero(as_tuple=True)[0] |
| 877 | slot_id = slot_mapping[idx] |
| 878 | block_idx = torch.div(slot_id, bs, rounding_mode="floor") |
| 879 | pos_in_block = slot_id % bs |
| 880 | |
| 881 | nope = kv_f32[idx, :NOPE_DIM].view(-1, num_quant_blocks, QUANT_BLOCK) |
| 882 | absmax = nope.abs().amax(dim=-1).clamp(min=1e-4) |
| 883 | exponent = torch.ceil(torch.log2(absmax / FP8_MAX)) |
| 884 | inv_scale = torch.exp2(-exponent) |
| 885 | scaled = (nope * inv_scale.unsqueeze(-1)).clamp(-FP8_MAX, FP8_MAX) |
| 886 | # Flush fp8-subnormal magnitudes to zero, matching the kernel's |
| 887 | # AVX512 cvt_fp32x16_to_fp8x16, which doesn't preserve exact |
| 888 | # subnormal encodings the way .to(torch.float8_e4m3fn) does. |
| 889 | fp8_min_normal = 2.0**-6 |
| 890 | scaled = torch.where(scaled.abs() < fp8_min_normal, 0.0, scaled) |
| 891 | nope_fp8_bytes = ( |
| 892 | scaled.to(torch.float8_e4m3fn).view(-1, NOPE_DIM).view(torch.uint8) |
| 893 | ) |
| 894 | scale_bytes = (exponent + 127.0).clamp(0, 255).to(torch.uint8) |
| 895 | rope_bf16_bytes = kv_bf16[idx, NOPE_DIM:].contiguous().view(torch.uint8) |
| 896 | |
| 897 | arange_nope = torch.arange(NOPE_DIM) |
| 898 | arange_rope = torch.arange(ROPE_DIM * 2) |
| 899 | arange_scale = torch.arange(num_quant_blocks) |
| 900 | row_idx = block_idx.unsqueeze(-1) |
| 901 | token_base = pos_in_block.unsqueeze(-1) * TOKEN_DATA_BYTES |
| 902 | |
| 903 | cache[row_idx, token_base + arange_nope] = nope_fp8_bytes |
| 904 | cache[row_idx, token_base + NOPE_DIM + arange_rope] = rope_bf16_bytes |
| 905 | scale_base = ( |
| 906 | bs * TOKEN_DATA_BYTES + pos_in_block.unsqueeze(-1) * SCALE_BYTES_PER_TOKEN |
| 907 | ) |
| 908 | cache[row_idx, scale_base + arange_scale] = scale_bytes |
| 909 | cache[block_idx, scale_base.squeeze(-1) + num_quant_blocks] = 0 |
| 910 | |
| 911 | return q_out |
| 912 | |
| 913 | |
| 914 | @pytest.mark.parametrize("num_tokens", [1, 64]) |
| 915 | @pytest.mark.parametrize("n_heads,padded_heads", [(1, 8), (8, 8), (8, 16), (64, 128)]) |
| 916 | def test_qnorm_rope_kv_insert_cpu_kernel_matches_eager( |
| 917 | num_tokens: int, n_heads: int, padded_heads: int |
| 918 | ): |
| 919 | torch.manual_seed(0) |
| 920 | dtype = torch.bfloat16 |
| 921 | eps = 1e-6 |
| 922 | max_pos = 4096 |
| 923 | block_size = 16 |
| 924 | |
| 925 | q = torch.randn(num_tokens, n_heads, HEAD_DIM, dtype=dtype) |
| 926 | kv = torch.randn(num_tokens, HEAD_DIM, dtype=dtype) |
| 927 | positions = torch.randint(0, max_pos, (num_tokens,), dtype=torch.int64) |
| 928 | cos_sin_cache = make_cos_sin_cache(max_pos, ROPE_DIM, torch.float32, "cpu") |
| 929 | |
| 930 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 931 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64) |
| 932 | if num_tokens > 1: |
| 933 | slot_mapping[-1] = -1 # exercise the "skip invalid slot" branch |
| 934 | |
| 935 | cache_ref = _make_cache(num_blocks, block_size) |
| 936 | cache_fused = cache_ref.clone() |
| 937 | |
| 938 | q_ref = _qnorm_kv_insert_eager( |
| 939 | q, |
| 940 | kv, |
| 941 | positions, |
| 942 | cache_ref, |
| 943 | slot_mapping, |
| 944 | cos_sin_cache, |
| 945 | padded_heads, |
| 946 | eps, |
| 947 | block_size, |
| 948 | ) |
| 949 | q_fused = fused_qnorm_rope_kv_insert_cpu( |
| 950 | q, |
| 951 | kv, |
| 952 | positions, |
| 953 | cache_fused, |
| 954 | slot_mapping, |
| 955 | cos_sin_cache, |
| 956 | padded_heads, |
| 957 | eps, |
| 958 | block_size, |
| 959 | ) |
| 960 | |
| 961 | torch.testing.assert_close(q_fused.float(), q_ref.float(), rtol=1e-2, atol=1e-2) |
| 962 | if n_heads < padded_heads: |
| 963 | pad_region = q_fused[:, n_heads:padded_heads] |
| 964 | assert pad_region.abs().max().item() == 0.0, ( |
| 965 | "padded head slots must be exact zero" |
| 966 | ) |
| 967 | |
| 968 | _assert_cache_parity(cache_fused, cache_ref, slot_mapping, block_size) |
| 969 | |
| 970 | |
| 971 | @pytest.mark.parametrize("pad", [1, 5]) |
| 972 | def test_qnorm_rope_kv_insert_cpu_kernel_dp_padding(pad: int): |
| 973 | """slot_mapping shorter than q/kv rows: the KV branch must only touch |
| 974 | the first num_tokens rows while Q-norm+RoPE still runs on all rows.""" |
| 975 | torch.manual_seed(1) |
| 976 | dtype = torch.bfloat16 |
| 977 | eps = 1e-6 |
| 978 | max_pos = 4096 |
| 979 | block_size = 16 |
| 980 | num_tokens = 17 |
| 981 | total = num_tokens + pad |
| 982 | padded_heads = 8 |
| 983 | |
| 984 | q = torch.randn(total, padded_heads, HEAD_DIM, dtype=dtype) |
| 985 | kv = torch.randn(total, HEAD_DIM, dtype=dtype) |
| 986 | positions = torch.randint(0, max_pos, (total,), dtype=torch.int64) |
| 987 | cos_sin_cache = make_cos_sin_cache(max_pos, ROPE_DIM, torch.float32, "cpu") |
| 988 | |
| 989 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 990 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64) |
| 991 | |
| 992 | cache_ref = _make_cache(num_blocks, block_size) |
| 993 | cache_fused = cache_ref.clone() |
| 994 | |
| 995 | q_ref = _qnorm_kv_insert_eager( |
| 996 | q, |
| 997 | kv, |
| 998 | positions, |
| 999 | cache_ref, |
| 1000 | slot_mapping, |
| 1001 | cos_sin_cache, |
| 1002 | padded_heads, |
| 1003 | eps, |
| 1004 | block_size, |
| 1005 | ) |
| 1006 | q_fused = fused_qnorm_rope_kv_insert_cpu( |
| 1007 | q, |
| 1008 | kv, |
| 1009 | positions, |
| 1010 | cache_fused, |
| 1011 | slot_mapping, |
| 1012 | cos_sin_cache, |
| 1013 | padded_heads, |
| 1014 | eps, |
| 1015 | block_size, |
| 1016 | ) |
| 1017 | |
| 1018 | torch.testing.assert_close(q_fused.float(), q_ref.float(), rtol=1e-2, atol=1e-2) |
| 1019 | _assert_cache_parity(cache_fused, cache_ref, slot_mapping, block_size) |
| 1020 | |
| 1021 | |
| 1022 | # --------------------------------------------------------------------------- |
| 1023 | # inverse_gptj_rope_o_proj_cpu (store_cache.cpp): DeepseekV4CPUAttention's |
| 1024 | # _o_proj de-rotation kernel |
| 1025 | # --------------------------------------------------------------------------- |
| 1026 | |
| 1027 | |
| 1028 | def _inverse_gptj_rope_o_proj_eager( |
| 1029 | o: torch.Tensor, |
| 1030 | positions: torch.Tensor, |
| 1031 | cos_sin_cache: torch.Tensor, |
| 1032 | rope_dim: int, |
| 1033 | ) -> torch.Tensor: |
| 1034 | """NoPE dims pass through as fp32; the RoPE segment is de-rotated by |
| 1035 | R(-pos) in GPT-J (interleaved even/odd) convention.""" |
| 1036 | nope_dim = o.shape[-1] - rope_dim |
| 1037 | half_dim = rope_dim // 2 |
| 1038 | |
| 1039 | o_f32 = o.float() |
| 1040 | out = torch.empty_like(o_f32) |
| 1041 | out[..., :nope_dim] = o_f32[..., :nope_dim] |
| 1042 | |
| 1043 | cos = cos_sin_cache[positions, :half_dim].unsqueeze(1) |
| 1044 | sin = cos_sin_cache[positions, half_dim:].unsqueeze(1) |
| 1045 | |
| 1046 | e = o_f32[..., nope_dim::2] |
| 1047 | od = o_f32[..., nope_dim + 1 :: 2] |
| 1048 | out[..., nope_dim::2] = e * cos + od * sin |
| 1049 | out[..., nope_dim + 1 :: 2] = od * cos - e * sin |
| 1050 | return out |
| 1051 | |
| 1052 | |
| 1053 | @pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16]) |
| 1054 | @pytest.mark.parametrize("num_tokens", [1, 17]) |
| 1055 | @pytest.mark.parametrize( |
| 1056 | "head_dim,rope_dim,num_heads", |
| 1057 | [ |
| 1058 | (512, 64, 8), # DeepseekV4CPUAttention._o_proj's actual shape |
| 1059 | (64, 64, 1), # nope_dim=0: no NoPE region at all |
| 1060 | ], |
| 1061 | ) |
| 1062 | def test_inverse_gptj_rope_o_proj_cpu_kernel_matches_eager( |
| 1063 | dtype: torch.dtype, |
| 1064 | num_tokens: int, |
| 1065 | head_dim: int, |
| 1066 | rope_dim: int, |
| 1067 | num_heads: int, |
| 1068 | ): |
| 1069 | torch.manual_seed(0) |
| 1070 | max_pos = 4096 |
| 1071 | |
| 1072 | o = torch.randn(num_tokens, num_heads, head_dim, dtype=dtype) |
| 1073 | positions = torch.randint(0, max_pos, (num_tokens,), dtype=torch.int64) |
| 1074 | cos_sin_cache = make_cos_sin_cache(max_pos, rope_dim, torch.float32, "cpu") |
| 1075 | |
| 1076 | out_ref = _inverse_gptj_rope_o_proj_eager(o, positions, cos_sin_cache, rope_dim) |
| 1077 | out_fused = inverse_gptj_rope_o_proj_cpu(o, positions, cos_sin_cache, rope_dim) |
| 1078 | |
| 1079 | torch.testing.assert_close(out_fused, out_ref, rtol=1e-2, atol=1e-2) |
| 1080 | |
| 1081 | |
| 1082 | # --------------------------------------------------------------------------- |
| 1083 | # fp8_paged_mqa_logits_cpu (paged_mqa_logits.cpp): sparse indexer's |
| 1084 | # decode-path MQA-logits kernel |
| 1085 | # --------------------------------------------------------------------------- |
| 1086 | |
| 1087 | |
| 1088 | def _rand_fp8(*shape: int) -> torch.Tensor: |
| 1089 | return torch.randn(shape, dtype=torch.float32).clamp(-3, 3).to(torch.float8_e4m3fn) |
| 1090 | |
| 1091 | |
| 1092 | def _pages_per_request(block_size: int, seq_lens: list[int]) -> list[int]: |
| 1093 | return [-(-seq_len // block_size) for seq_len in seq_lens] |
| 1094 | |
| 1095 | |
| 1096 | def _write_page( |
| 1097 | page_row: torch.Tensor, block_size: int, k_fp8: torch.Tensor, scale: torch.Tensor |
| 1098 | ) -> None: |
| 1099 | """Write k_fp8 ([block_size, INDEXER_HEAD_DIM]) and scale ([block_size] |
| 1100 | fp32) into one page row: K-region then scale-region, matching |
| 1101 | compress_norm_rope_store_indexer_cpu's layout.""" |
| 1102 | k_region = page_row[: block_size * INDEXER_HEAD_DIM] |
| 1103 | k_region.copy_(k_fp8.view(torch.uint8).flatten()) |
| 1104 | scale_region = page_row[block_size * INDEXER_HEAD_DIM :].view(torch.float32) |
| 1105 | scale_region[:].copy_(scale) |
| 1106 | |
| 1107 | |
| 1108 | def _reference_paged_mqa_logits( |
| 1109 | q_fp8: torch.Tensor, |
| 1110 | weight: torch.Tensor, |
| 1111 | ref_k: list[torch.Tensor], |
| 1112 | ref_scale: list[torch.Tensor], |
| 1113 | max_seq_len: int, |
| 1114 | ) -> torch.Tensor: |
| 1115 | batch_size, _num_heads, _ = q_fp8.shape |
| 1116 | out = torch.zeros((batch_size, max_seq_len), dtype=torch.float32) |
| 1117 | for b in range(batch_size): |
| 1118 | k = ref_k[b].float() # [seq_len, head_dim] |
| 1119 | scale = ref_scale[b] # [seq_len] |
| 1120 | q = q_fp8[b].float() # [heads, head_dim] |
| 1121 | w = weight[b] # [heads] |
| 1122 | dot = torch.einsum("jd,hd->jh", k, q).clamp(min=0.0) |
| 1123 | score = (dot * w.unsqueeze(0)).sum(dim=1) |
| 1124 | out[b, : k.shape[0]] = score * scale |
| 1125 | return out |
| 1126 | |
| 1127 | |
| 1128 | def _build_paged_mqa_logits_batch( |
| 1129 | block_size: int, |
| 1130 | seq_lens: list[int], |
| 1131 | num_heads: int, |
| 1132 | pages: torch.Tensor, |
| 1133 | seed: int, |
| 1134 | ): |
| 1135 | """Fill pages (a pre-sized [n_pages, buf_width] uint8 tensor, possibly |
| 1136 | non-contiguous in dim 0) with deterministic-per-seed random K/scale |
| 1137 | data.""" |
| 1138 | torch.manual_seed(seed) |
| 1139 | batch_size = len(seq_lens) |
| 1140 | max_seq_len = max(seq_lens) |
| 1141 | per_request = _pages_per_request(block_size, seq_lens) |
| 1142 | pages_per_batch = max(per_request) |
| 1143 | page_table = torch.full((batch_size, pages_per_batch), -1, dtype=torch.int32) |
| 1144 | |
| 1145 | ref_k: list[torch.Tensor] = [] |
| 1146 | ref_scale: list[torch.Tensor] = [] |
| 1147 | next_page = 0 |
| 1148 | for b, seq_len in enumerate(seq_lens): |
| 1149 | k_chunks = [] |
| 1150 | scale_chunks = [] |
| 1151 | for p in range(per_request[b]): |
| 1152 | page_table[b, p] = next_page |
| 1153 | k = _rand_fp8(block_size, INDEXER_HEAD_DIM) |
| 1154 | scale = torch.rand(block_size, dtype=torch.float32) * 0.05 + 0.01 |
| 1155 | _write_page(pages[next_page], block_size, k, scale) |
| 1156 | tokens_here = min(block_size, seq_len - p * block_size) |
| 1157 | k_chunks.append(k[:tokens_here]) |
| 1158 | scale_chunks.append(scale[:tokens_here]) |
| 1159 | next_page += 1 |
| 1160 | ref_k.append(torch.cat(k_chunks, dim=0)) |
| 1161 | ref_scale.append(torch.cat(scale_chunks, dim=0)) |
| 1162 | |
| 1163 | q_fp8 = _rand_fp8(batch_size, num_heads, INDEXER_HEAD_DIM) |
| 1164 | weight = torch.rand(batch_size, num_heads, dtype=torch.float32) * 0.1 |
| 1165 | seq_lens_t = torch.tensor(seq_lens, dtype=torch.int32) |
| 1166 | return q_fp8, weight, seq_lens_t, page_table, ref_k, ref_scale, max_seq_len |
| 1167 | |
| 1168 | |
| 1169 | def _paged_mqa_logits_page_layout( |
| 1170 | block_size: int, seq_lens: list[int] |
| 1171 | ) -> tuple[int, int]: |
| 1172 | n_pages = sum(_pages_per_request(block_size, seq_lens)) |
| 1173 | buf_width = block_size * INDEXER_HEAD_DIM + block_size * INDEXER_SCALE_BYTES |
| 1174 | return n_pages, buf_width |
| 1175 | |
| 1176 | |
| 1177 | @pytest.mark.parametrize("block_size", [64, 256]) |
| 1178 | def test_fp8_paged_mqa_logits_cpu_matches_eager(block_size: int): |
| 1179 | num_heads = 4 |
| 1180 | seq_lens = [1, block_size + 7, 3 * block_size] |
| 1181 | n_pages, buf_width = _paged_mqa_logits_page_layout(block_size, seq_lens) |
| 1182 | pages = torch.zeros((n_pages, buf_width), dtype=torch.uint8) |
| 1183 | |
| 1184 | q_fp8, weight, seq_lens_t, page_table, ref_k, ref_scale, max_seq_len = ( |
| 1185 | _build_paged_mqa_logits_batch(block_size, seq_lens, num_heads, pages, seed=0) |
| 1186 | ) |
| 1187 | |
| 1188 | logits = fp8_paged_mqa_logits_cpu( |
| 1189 | q_fp8, pages, weight, seq_lens_t, page_table, block_size, max_seq_len |
| 1190 | ) |
| 1191 | ref = _reference_paged_mqa_logits(q_fp8, weight, ref_k, ref_scale, max_seq_len) |
| 1192 | |
| 1193 | for b, seq_len in enumerate(seq_lens): |
| 1194 | torch.testing.assert_close( |
| 1195 | logits[b, :seq_len], ref[b, :seq_len], atol=1e-2, rtol=1e-2 |
| 1196 | ) |
| 1197 | |
| 1198 | |
| 1199 | def test_fp8_paged_mqa_logits_cpu_non_contiguous_kv_cache_row_stride(): |
| 1200 | """The paged K-cache in production is a per-layer view into a shared |
| 1201 | multi-layer allocation, so its page stride generally exceeds the page |
| 1202 | byte width -- check the kernel reads a padded/strided buffer identically |
| 1203 | to an equivalent tightly-packed one.""" |
| 1204 | block_size = 64 |
| 1205 | num_heads = 4 |
| 1206 | seq_lens = [1, block_size + 7, 3 * block_size] |
| 1207 | n_pages, buf_width = _paged_mqa_logits_page_layout(block_size, seq_lens) |
| 1208 | cache_pad = 48 |
| 1209 | |
| 1210 | tight_pages = torch.zeros((n_pages, buf_width), dtype=torch.uint8) |
| 1211 | q_fp8, weight, seq_lens_t, page_table, _, _, max_seq_len = ( |
| 1212 | _build_paged_mqa_logits_batch( |
| 1213 | block_size, seq_lens, num_heads, tight_pages, seed=0 |
| 1214 | ) |
| 1215 | ) |
| 1216 | |
| 1217 | padded = torch.zeros((n_pages, buf_width + cache_pad), dtype=torch.uint8) |
| 1218 | strided_pages = padded[:, :buf_width] |
| 1219 | _build_paged_mqa_logits_batch( |
| 1220 | block_size, seq_lens, num_heads, strided_pages, seed=0 |
| 1221 | ) |
| 1222 | assert strided_pages.stride(0) != strided_pages.size(1) |
| 1223 | |
| 1224 | logits_tight = fp8_paged_mqa_logits_cpu( |
| 1225 | q_fp8, tight_pages, weight, seq_lens_t, page_table, block_size, max_seq_len |
| 1226 | ) |
| 1227 | logits_strided = fp8_paged_mqa_logits_cpu( |
| 1228 | q_fp8, strided_pages, weight, seq_lens_t, page_table, block_size, max_seq_len |
| 1229 | ) |
| 1230 | # Only [:seq_len] is ever written (and ever read downstream by |
| 1231 | # topk_transform_512_cpu) -- positions beyond that are uninitialized, so |
| 1232 | # comparing the full row would be comparing unrelated garbage. |
| 1233 | for b, seq_len in enumerate(seq_lens): |
| 1234 | torch.testing.assert_close( |
| 1235 | logits_tight[b, :seq_len], logits_strided[b, :seq_len] |
| 1236 | ) |
| 1237 | |
| 1238 | |
| 1239 | # --------------------------------------------------------------------------- |
| 1240 | # fused_indexer_q_rope_quant_cpu (indexer.cpp): indexer Q-side RoPE + FP8 |
| 1241 | # quant, and the dispatcher wiring that routes to it on CPU |
| 1242 | # --------------------------------------------------------------------------- |
| 1243 | # |
| 1244 | # Independent eager reference: this kernel's absmax reduction is over an |
| 1245 | # unusual mix (raw NoPE values plus a bf16-round-tripped RoPE half), which |
| 1246 | # per_token_group_quant_fp8 does not replicate. Allow <=1 fp8 ULP: the RoPE |
| 1247 | # halves round a value sitting exactly on a tie to either side. |
| 1248 | |
| 1249 | INDEXER_Q_N_HEAD = 64 |
| 1250 | INDEXER_Q_MAX_POS = 4096 |
| 1251 | |
| 1252 | |
| 1253 | def _indexer_q_rope_quant_reference( |
| 1254 | positions: torch.Tensor, |
| 1255 | q: torch.Tensor, |
| 1256 | cos_sin_cache: torch.Tensor, |
| 1257 | weights: torch.Tensor, |
| 1258 | softmax_scale: float, |
| 1259 | head_scale: float, |
| 1260 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 1261 | _num_tokens, _num_heads, head_dim = q.shape |
| 1262 | half_rot_dim = cos_sin_cache.shape[-1] // 2 |
| 1263 | rot_dim = 2 * half_rot_dim |
| 1264 | nope_dim = head_dim - rot_dim |
| 1265 | |
| 1266 | q_f32 = q.float() |
| 1267 | cos_sin_f32 = cos_sin_cache.float() |
| 1268 | pos_cos_sin = cos_sin_f32[positions] # [T, rot_dim] |
| 1269 | cos = pos_cos_sin[:, :half_rot_dim].unsqueeze(1) |
| 1270 | sin = pos_cos_sin[:, half_rot_dim:].unsqueeze(1) |
| 1271 | |
| 1272 | nope = q_f32[..., :nope_dim] |
| 1273 | rot = q_f32[..., nope_dim:] |
| 1274 | x_even = rot[..., 0::2] |
| 1275 | x_odd = rot[..., 1::2] |
| 1276 | r_even = (x_even * cos - x_odd * sin).to(torch.bfloat16).to(torch.float32) |
| 1277 | r_odd = (x_odd * cos + x_even * sin).to(torch.bfloat16).to(torch.float32) |
| 1278 | rope_out = torch.stack((r_even, r_odd), dim=-1).flatten(-2) |
| 1279 | |
| 1280 | combined = torch.cat((nope, rope_out), dim=-1) # [T, H, head_dim] |
| 1281 | amax = combined.abs().amax(dim=-1).clamp(min=1e-4) # [T, H] |
| 1282 | exponent = torch.ceil(torch.log2(amax / FP8_MAX)) |
| 1283 | scale = torch.exp2(exponent) |
| 1284 | inv_scale = scale.reciprocal().unsqueeze(-1) |
| 1285 | q_fp8 = (combined * inv_scale).clamp(-FP8_MAX, FP8_MAX).to(torch.float8_e4m3fn) |
| 1286 | |
| 1287 | weights_out = weights.to(torch.float32) * scale * softmax_scale * head_scale |
| 1288 | return q_fp8, weights_out |
| 1289 | |
| 1290 | |
| 1291 | @pytest.mark.parametrize("num_tokens", [1, 257]) |
| 1292 | def test_fused_indexer_q_rope_quant_cpu_matches_eager(num_tokens: int): |
| 1293 | torch.manual_seed(0) |
| 1294 | |
| 1295 | q = torch.randn(num_tokens, INDEXER_Q_N_HEAD, INDEXER_HEAD_DIM, dtype=torch.float32) |
| 1296 | positions = torch.randint(0, INDEXER_Q_MAX_POS, (num_tokens,), dtype=torch.int64) |
| 1297 | cos_sin_cache = torch.randn(INDEXER_Q_MAX_POS, ROPE_DIM, dtype=torch.float32) |
| 1298 | weights = torch.randn(num_tokens, INDEXER_Q_N_HEAD, dtype=torch.float32) |
| 1299 | softmax_scale = INDEXER_HEAD_DIM**-0.5 |
| 1300 | head_scale = INDEXER_Q_N_HEAD**-0.5 |
| 1301 | |
| 1302 | q_fp8_ref, weights_ref = _indexer_q_rope_quant_reference( |
| 1303 | positions, q, cos_sin_cache, weights, softmax_scale, head_scale |
| 1304 | ) |
| 1305 | |
| 1306 | q_fp8_fused = torch.empty_like(q, dtype=torch.float8_e4m3fn) |
| 1307 | weights_fused = torch.empty_like(weights, dtype=torch.float32) |
| 1308 | fused_indexer_q_rope_quant_cpu( |
| 1309 | positions, |
| 1310 | q, |
| 1311 | cos_sin_cache, |
| 1312 | q_fp8_fused, |
| 1313 | weights, |
| 1314 | softmax_scale, |
| 1315 | head_scale, |
| 1316 | weights_fused, |
| 1317 | ) |
| 1318 | |
| 1319 | max_ulp = int(fp8_ulp_distance(q_fp8_fused, q_fp8_ref).max().item()) |
| 1320 | assert max_ulp <= 1, f"q_fp8 differs by {max_ulp} ULP (>1)" |
| 1321 | torch.testing.assert_close(weights_fused, weights_ref, rtol=1e-5, atol=1e-6) |
| 1322 | |
| 1323 | |
| 1324 | def test_fused_indexer_q_rope_quant_dispatches_to_cpu_kernel(): |
| 1325 | """End-to-end wiring check: the production dispatcher must route to the |
| 1326 | CPU op on a CPU tensor, not silently fall through to the Triton kernel |
| 1327 | (which would also "work" under triton-cpu, masking a dispatch |
| 1328 | regression). The dispatcher does its CPU import inside the CPU branch, |
| 1329 | so the patch target is the source module.""" |
| 1330 | torch.manual_seed(0) |
| 1331 | num_tokens = 1 |
| 1332 | |
| 1333 | q = torch.randn(num_tokens, INDEXER_Q_N_HEAD, INDEXER_HEAD_DIM, dtype=torch.float32) |
| 1334 | positions = torch.randint(0, INDEXER_Q_MAX_POS, (num_tokens,), dtype=torch.int64) |
| 1335 | cos_sin_cache = torch.randn(INDEXER_Q_MAX_POS, ROPE_DIM, dtype=torch.float32) |
| 1336 | weights = torch.randn(num_tokens, INDEXER_Q_N_HEAD, dtype=torch.float32) |
| 1337 | softmax_scale = INDEXER_HEAD_DIM**-0.5 |
| 1338 | head_scale = INDEXER_Q_N_HEAD**-0.5 |
| 1339 | |
| 1340 | q_fp8_ref, weights_ref = _indexer_q_rope_quant_reference( |
| 1341 | positions, q, cos_sin_cache, weights, softmax_scale, head_scale |
| 1342 | ) |
| 1343 | |
| 1344 | with mock.patch("vllm._custom_ops.fused_indexer_q_rope_quant_cpu") as mocked: |
| 1345 | fused_indexer_q_rope_quant( |
| 1346 | positions, |
| 1347 | q.clone(), |
| 1348 | cos_sin_cache, |
| 1349 | weights, |
| 1350 | softmax_scale, |
| 1351 | head_scale, |
| 1352 | use_fp4=False, |
| 1353 | ) |
| 1354 | assert mocked.called, ( |
| 1355 | "fused_indexer_q_rope_quant did not dispatch to the CPU kernel " |
| 1356 | "on a CPU tensor" |
| 1357 | ) |
| 1358 | |
| 1359 | q_fp8_fused, weights_fused = fused_indexer_q_rope_quant( |
| 1360 | positions, |
| 1361 | q.clone(), |
| 1362 | cos_sin_cache, |
| 1363 | weights, |
| 1364 | softmax_scale, |
| 1365 | head_scale, |
| 1366 | use_fp4=False, |
| 1367 | ) |
| 1368 | max_ulp = int(fp8_ulp_distance(q_fp8_fused, q_fp8_ref).max().item()) |
| 1369 | assert max_ulp <= 1, f"q_fp8 differs by {max_ulp} ULP (>1)" |
| 1370 | torch.testing.assert_close(weights_fused, weights_ref, rtol=1e-5, atol=1e-6) |
| 1371 | |
| 1372 | |
| 1373 | # --------------------------------------------------------------------------- |
| 1374 | # mhc_pre_cpu / mhc_post_cpu / hc_head_fused_cpu (mhc.cpp): CPU ports of the |
| 1375 | # mHC gating kernels |
| 1376 | # --------------------------------------------------------------------------- |
| 1377 | |
| 1378 | |
| 1379 | def _rmsnorm_nw_torch(x: torch.Tensor, eps: float) -> torch.Tensor: |
| 1380 | """Weight-free RMSNorm over the last dim, matching hc_head_fused_cpu.""" |
| 1381 | x_f32 = x.float() |
| 1382 | var = x_f32.square().mean(dim=-1, keepdim=True) |
| 1383 | return (x_f32 * torch.rsqrt(var + eps)).to(x.dtype) |
| 1384 | |
| 1385 | |
| 1386 | def _hc_head_torch( |
| 1387 | hidden_states: torch.Tensor, |
| 1388 | hc_fn: torch.Tensor, |
| 1389 | hc_scale: torch.Tensor, |
| 1390 | hc_base: torch.Tensor, |
| 1391 | rms_norm_eps: float, |
| 1392 | hc_eps: float, |
| 1393 | ) -> torch.Tensor: |
| 1394 | """Eager reference for hc_head_fused_cpu: round-trips the RMSNorm |
| 1395 | output through bf16 before the linear, matching the kernel's bf16 |
| 1396 | arithmetic.""" |
| 1397 | hc_mult, hidden_size = hidden_states.shape[-2:] |
| 1398 | outer_shape = hidden_states.shape[:-2] |
| 1399 | hs_flat = hidden_states.reshape(-1, hc_mult, hidden_size) |
| 1400 | |
| 1401 | x_flat = hs_flat.flatten(-2) |
| 1402 | x_normed = _rmsnorm_nw_torch(x_flat, rms_norm_eps) |
| 1403 | mixes = torch.nn.functional.linear(x_normed.float(), hc_fn) |
| 1404 | pre = torch.sigmoid(mixes * hc_scale + hc_base) + hc_eps |
| 1405 | |
| 1406 | out = torch.sum(pre.unsqueeze(-1) * hs_flat.float(), dim=1).to(hidden_states.dtype) |
| 1407 | return out.view(*outer_shape, hidden_size) |
| 1408 | |
| 1409 | |
| 1410 | @pytest.mark.parametrize("num_tokens", [1, 128]) |
| 1411 | @pytest.mark.parametrize("hidden_size", [4096, 7168]) |
| 1412 | def test_mhc_pre_cpu(num_tokens: int, hidden_size: int): |
| 1413 | """mhc_pre_cpu vs. the mhc_pre_torch eager reference it replaces.""" |
| 1414 | torch.set_default_device(DEVICE) |
| 1415 | set_random_seed(0) |
| 1416 | hc_mult = 4 |
| 1417 | |
| 1418 | hc_mult3 = 2 * hc_mult + hc_mult * hc_mult |
| 1419 | residual = torch.randn((num_tokens, hc_mult, hidden_size), dtype=torch.bfloat16) |
| 1420 | fn = torch.randn((hc_mult3, hc_mult * hidden_size), dtype=torch.float32) * 1e-4 |
| 1421 | hc_scale = torch.randn((3,), dtype=torch.float32) * 0.1 |
| 1422 | hc_base = torch.randn((hc_mult3,), dtype=torch.float32) * 0.1 |
| 1423 | |
| 1424 | rms_eps = hc_eps = 1e-6 |
| 1425 | hc_post_mult_value = 2.0 |
| 1426 | sinkhorn_repeat = 8 |
| 1427 | |
| 1428 | ref = mhc_kernels.mhc_pre_torch( |
| 1429 | residual, |
| 1430 | fn, |
| 1431 | hc_scale, |
| 1432 | hc_base, |
| 1433 | rms_eps, |
| 1434 | hc_eps, |
| 1435 | hc_eps, |
| 1436 | hc_post_mult_value, |
| 1437 | sinkhorn_repeat, |
| 1438 | ) |
| 1439 | out = mhc_kernels.mhc_pre_cpu( |
| 1440 | residual, |
| 1441 | fn, |
| 1442 | hc_scale, |
| 1443 | hc_base, |
| 1444 | rms_eps, |
| 1445 | hc_eps, |
| 1446 | hc_eps, |
| 1447 | hc_post_mult_value, |
| 1448 | sinkhorn_repeat, |
| 1449 | ) |
| 1450 | |
| 1451 | for actual, expected in zip(out, ref, strict=True): |
| 1452 | torch.testing.assert_close(actual, expected, atol=5e-2, rtol=1e-2) |
| 1453 | |
| 1454 | |
| 1455 | @pytest.mark.parametrize("num_tokens", [1, 128]) |
| 1456 | @pytest.mark.parametrize("hidden_size", [4096, 7168]) |
| 1457 | def test_mhc_post_cpu(num_tokens: int, hidden_size: int): |
| 1458 | """mhc_post_cpu vs. the mhc_post_torch eager reference it replaces.""" |
| 1459 | torch.set_default_device(DEVICE) |
| 1460 | set_random_seed(0) |
| 1461 | hc_mult = 4 |
| 1462 | |
| 1463 | x = torch.randn((num_tokens, hidden_size), dtype=torch.bfloat16) |
| 1464 | residual = torch.randn((num_tokens, hc_mult, hidden_size), dtype=torch.bfloat16) |
| 1465 | post_layer_mix = torch.randn((num_tokens, hc_mult, 1), dtype=torch.float32) |
| 1466 | comb_res_mix = torch.randn((num_tokens, hc_mult, hc_mult), dtype=torch.float32) |
| 1467 | |
| 1468 | ref = mhc_kernels.mhc_post_torch(x, residual, post_layer_mix, comb_res_mix) |
| 1469 | out = mhc_kernels.mhc_post_cpu(x, residual, post_layer_mix, comb_res_mix) |
| 1470 | |
| 1471 | torch.testing.assert_close(out, ref, atol=5e-2, rtol=1e-2) |
| 1472 | |
| 1473 | |
| 1474 | @pytest.mark.parametrize("num_tokens", [1, 128]) |
| 1475 | @pytest.mark.parametrize("hidden_size", [4096, 7168]) |
| 1476 | def test_hc_head_cpu(num_tokens: int, hidden_size: int): |
| 1477 | """hc_head_fused_cpu vs. the _hc_head_torch eager reference it |
| 1478 | replaces.""" |
| 1479 | torch.set_default_device(DEVICE) |
| 1480 | set_random_seed(0) |
| 1481 | hc_mult = 4 |
| 1482 | |
| 1483 | residual = torch.randn((num_tokens, hc_mult, hidden_size), dtype=torch.bfloat16) |
| 1484 | fn = torch.randn((hc_mult, hc_mult * hidden_size), dtype=torch.float32) * 1e-4 |
| 1485 | hc_scale = torch.randn((1,), dtype=torch.float32) * 0.1 |
| 1486 | hc_base = torch.randn((hc_mult,), dtype=torch.float32) * 0.1 |
| 1487 | rms_eps = hc_eps = 1e-6 |
| 1488 | |
| 1489 | ref = _hc_head_torch(residual, fn, hc_scale, hc_base, rms_eps, hc_eps) |
| 1490 | out = mhc_kernels.hc_head_fused_cpu( |
| 1491 | residual, fn, hc_scale, hc_base, rms_eps, hc_eps |
| 1492 | ) |
| 1493 | |
| 1494 | torch.testing.assert_close(out, ref, atol=5e-2, rtol=1e-2) |