| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Round-trip tests for compressor → FP8 quant + KV cache insert → gather + dequant. |
| 4 | |
| 5 | These tests cover: |
| 6 | A) DeepseekV4 Attention: head_dim=512 (448 FP8 nope + 64 bf16 rope), quant_block=64 |
| 7 | B) Fused dequant+gather K cache |
| 8 | C) Indexer: head_dim=128 (all FP8), quant_block=128 |
| 9 | D) DeepseekV4 Attention magnitude range: correctness across small/large values |
| 10 | E) Indexer fused Triton kernel: compress+norm+rope+quant+insert |
| 11 | F) Indexer fused two-stage Triton kernel: head=512 cr>=128 (no-overlap) |
| 12 | """ |
| 13 | |
| 14 | import math |
| 15 | from types import SimpleNamespace |
| 16 | |
| 17 | import pytest |
| 18 | import torch |
| 19 | |
| 20 | from vllm import _custom_ops as ops |
| 21 | from vllm.models.deepseek_v4.common.ops import ( |
| 22 | dequantize_and_gather_k_cache, |
| 23 | quantize_and_insert_k_cache, |
| 24 | ) |
| 25 | from vllm.models.deepseek_v4.common.ops.fused_compress_quant_cache import ( |
| 26 | _fused_kv_compress_norm_rope_insert_indexer_attn, |
| 27 | _fused_kv_compress_norm_rope_insert_indexer_mxfp4_attn, |
| 28 | _launch_two_stage_sparse_attn_compressor, |
| 29 | compress_norm_rope_store_triton, |
| 30 | ) |
| 31 | from vllm.models.deepseek_v4.compressor import _get_c128_boundary |
| 32 | from vllm.platforms import current_platform |
| 33 | from vllm.v1.attention.backends.mla.compressor_utils import ( |
| 34 | get_dspark_swa_index_width, |
| 35 | ) |
| 36 | from vllm.v1.attention.ops.rocm_aiter_mla_sparse import ( |
| 37 | cp_gather_indexer_k_quant_cache_triton, |
| 38 | indexer_k_quant_and_cache_triton, |
| 39 | ) |
| 40 | |
| 41 | from .test_fused_indexer_q_rope_quant import quantize_to_mxfp4 |
| 42 | |
| 43 | |
| 44 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 45 | @pytest.mark.parametrize( |
| 46 | "cache_dtype,kv_mxfp8", |
| 47 | [ |
| 48 | (torch.uint8, False), |
| 49 | (torch.uint8, True), |
| 50 | (torch.bfloat16, False), |
| 51 | (torch.float8_e4m3fn, False), |
| 52 | ], |
| 53 | ) |
| 54 | @pytest.mark.parametrize("num_tokens", [1, 17, 1023, 1024]) |
| 55 | @pytest.mark.parametrize("use_graph", [False, True]) |
| 56 | def test_dspark_context_kv_matches_query_insert( |
| 57 | cache_dtype, kv_mxfp8, num_tokens, use_graph |
| 58 | ): |
| 59 | """KV-only insertion must preserve every cache byte, including graph replay.""" |
| 60 | from vllm.models.deepseek_v41.nvidia.dspark import _insert_context_kv |
| 61 | |
| 62 | torch.manual_seed(42) |
| 63 | block_size = 256 |
| 64 | num_blocks = math.ceil((num_tokens + 7) / block_size) |
| 65 | row_size = (528 if kv_mxfp8 else 584) if cache_dtype == torch.uint8 else 512 |
| 66 | page_stride = math.ceil(block_size * row_size / 576) * 576 + 576 |
| 67 | backing = torch.full((num_blocks, page_stride), 3, device="cuda").to(cache_dtype) |
| 68 | cache = backing.as_strided( |
| 69 | (num_blocks, block_size, row_size), (page_stride, row_size, 1) |
| 70 | ) |
| 71 | reference_backing = backing.clone() |
| 72 | reference = reference_backing.as_strided(cache.shape, cache.stride()) |
| 73 | kv = torch.randn(num_tokens + 3, 512, device="cuda", dtype=torch.bfloat16) |
| 74 | positions = torch.arange(num_tokens + 3, device="cuda") + 7 |
| 75 | angles = torch.randn(num_tokens + 16, 32, device="cuda") |
| 76 | cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1) |
| 77 | slots = torch.randperm(num_blocks * block_size, device="cuda")[:num_tokens] |
| 78 | slots[::5] = -1 |
| 79 | scale = torch.tensor([0.7], device="cuda") |
| 80 | # Only what _insert_context_kv actually reads: the record width comes off |
| 81 | # the cache tensor, so no per-record width attribute is stubbed here. |
| 82 | attn = SimpleNamespace( |
| 83 | swa_cache_layer=SimpleNamespace(kv_cache=cache, block_size=block_size), |
| 84 | kv_mxfp8=kv_mxfp8, |
| 85 | rotary_emb=SimpleNamespace(cos_sin_cache=cos_sin), |
| 86 | _flashinfer_fp8_kv_scale=scale, |
| 87 | ) |
| 88 | |
| 89 | def legacy_insert(): |
| 90 | q = torch.zeros(kv.shape[0], 8, 512, dtype=kv.dtype, device="cuda") |
| 91 | if cache_dtype == torch.uint8: |
| 92 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert( |
| 93 | q, |
| 94 | kv, |
| 95 | reference.view(num_blocks, -1), |
| 96 | slots, |
| 97 | positions, |
| 98 | cos_sin, |
| 99 | 8, |
| 100 | 1e-20, |
| 101 | block_size, |
| 102 | True, |
| 103 | kv_mxfp8, |
| 104 | ) |
| 105 | elif cache_dtype == torch.bfloat16: |
| 106 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert( |
| 107 | q, |
| 108 | kv, |
| 109 | reference, |
| 110 | slots, |
| 111 | positions, |
| 112 | cos_sin, |
| 113 | 1e-20, |
| 114 | block_size, |
| 115 | ) |
| 116 | else: |
| 117 | q_fp8 = torch.empty_like(q, dtype=cache_dtype) |
| 118 | torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert( |
| 119 | q, |
| 120 | kv, |
| 121 | q_fp8, |
| 122 | reference, |
| 123 | slots, |
| 124 | positions, |
| 125 | cos_sin, |
| 126 | scale, |
| 127 | scale, |
| 128 | 1e-20, |
| 129 | block_size, |
| 130 | ) |
| 131 | |
| 132 | def insert(): |
| 133 | _insert_context_kv(attn, kv, positions, slots) |
| 134 | |
| 135 | insert() |
| 136 | graph = None |
| 137 | if use_graph: |
| 138 | graph = torch.cuda.CUDAGraph() |
| 139 | with torch.cuda.graph(graph): |
| 140 | insert() |
| 141 | for iteration in range(3): |
| 142 | kv.normal_() |
| 143 | positions.add_(1) |
| 144 | if iteration == 2: |
| 145 | slots.fill_(-1) |
| 146 | backing.view(torch.uint8).fill_(165) |
| 147 | reference_backing.copy_(backing) |
| 148 | before = kv.clone() |
| 149 | legacy_insert() |
| 150 | insert() if graph is None else graph.replay() |
| 151 | torch.testing.assert_close( |
| 152 | backing.view(torch.uint8), |
| 153 | reference_backing.view(torch.uint8), |
| 154 | rtol=0, |
| 155 | atol=0, |
| 156 | ) |
| 157 | torch.testing.assert_close(kv, before, rtol=0, atol=0) |
| 158 | |
| 159 | |
| 160 | @pytest.mark.skipif( |
| 161 | not current_platform.is_cuda_alike(), reason="graph capture coverage" |
| 162 | ) |
| 163 | @pytest.mark.parametrize("compress_ratio", [1, 2]) |
| 164 | @pytest.mark.parametrize("use_graph", [False, True]) |
| 165 | @pytest.mark.parametrize( |
| 166 | "lengths", |
| 167 | [(1, 5, 13), (1, 1, 1), (1, 2, 0), (7, 8, 8), (2, 3, 3)], |
| 168 | ids=["prefill", "decode", "empty", "wrap_threshold", "packed_boundary"], |
| 169 | ) |
| 170 | def test_v41_fused_save_compress_and_insert( |
| 171 | compress_ratio: int, use_graph: bool, lengths: tuple[int, ...] |
| 172 | ): |
| 173 | """Ring states: a group boundary whose previous token is outside the chunk |
| 174 | pools the ring row that chunk left behind, even when this chunk overwrites |
| 175 | that row; only the last ``capacity`` tokens of a chunk are stored. |
| 176 | |
| 177 | Exercise historical group starts, adjacent requests, a chunk longer than |
| 178 | the ring, single-token decode pairs, empty requests, wrap thresholds at |
| 179 | odd and even starts, padded graph rows, and padded physical pages against |
| 180 | a Torch reference with BF16 rounding before RoPE. Replays change both |
| 181 | inputs and history so stale scratch/cache values cannot satisfy the test. |
| 182 | """ |
| 183 | from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import ( |
| 184 | fused_save_compress_norm, |
| 185 | rope_quant_insert, |
| 186 | ) |
| 187 | |
| 188 | torch.manual_seed(42) |
| 189 | device = "cuda" |
| 190 | capacity = 8 |
| 191 | starts = [7, 8, 13] |
| 192 | actual = sum(lengths) |
| 193 | num_tokens = actual + 5 |
| 194 | positions = torch.zeros(num_tokens, dtype=torch.int64, device=device) |
| 195 | req_ids = torch.zeros(actual + 2, dtype=torch.int32, device=device) |
| 196 | state_slots = torch.full((actual + 2,), -1, dtype=torch.int64, device=device) |
| 197 | cache_slots = torch.full_like(state_slots, -1) |
| 198 | query_start_loc = torch.tensor( |
| 199 | [0, *torch.tensor(lengths).cumsum(0).tolist()], dtype=torch.int32 |
| 200 | ).to(device) |
| 201 | ring_blocks = torch.randperm(6, device=device)[:3] |
| 202 | raw = torch.randn(num_tokens, 512 * compress_ratio, device=device) |
| 203 | norm = torch.randn(512, dtype=torch.bfloat16, device=device) |
| 204 | latent = torch.empty(num_tokens, 512, dtype=torch.bfloat16, device=device) |
| 205 | angles = torch.randn(32, 32, device=device) |
| 206 | cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1) |
| 207 | # capacity x 1024 floats rounded up to 576-byte alignment. |
| 208 | state_backing = torch.empty(6, 8208, device=device) |
| 209 | state = state_backing.as_strided((6, capacity, 1024), (8208, 1024, 1)) |
| 210 | cache_block = 256 // compress_ratio |
| 211 | cache_stride = math.ceil(cache_block * 584 / 576) * 576 |
| 212 | cache_backing = torch.empty(3, cache_stride, dtype=torch.uint8, device=device) |
| 213 | cache = cache_backing.as_strided((3, cache_block, 584), (cache_stride, 584, 1)) |
| 214 | cursor = 0 |
| 215 | for req, (start, length) in enumerate(zip(starts, lengths)): |
| 216 | pos = torch.arange(start, start + length, device=device) |
| 217 | rows = slice(cursor, cursor + length) |
| 218 | positions[rows] = pos |
| 219 | req_ids[rows] = req |
| 220 | state_slots[rows] = ring_blocks[req] * capacity + pos % capacity |
| 221 | cache_slots[rows] = (2 - req) * cache_block + pos // compress_ratio |
| 222 | cursor += length |
| 223 | |
| 224 | has_ring = compress_ratio == 2 |
| 225 | |
| 226 | def run(): |
| 227 | fused_save_compress_norm( |
| 228 | raw, |
| 229 | positions, |
| 230 | state if has_ring else None, |
| 231 | state_slots, |
| 232 | query_start_loc if has_ring else None, |
| 233 | req_ids if has_ring else None, |
| 234 | norm, |
| 235 | 1e-20, |
| 236 | compress_ratio, |
| 237 | latent, |
| 238 | ) |
| 239 | rope_quant_insert( |
| 240 | latent, positions, cos_sin, cache, cache_slots, compress_ratio |
| 241 | ) |
| 242 | |
| 243 | state_backing.normal_() |
| 244 | run() |
| 245 | graph = None |
| 246 | if use_graph: |
| 247 | stream = torch.cuda.Stream() |
| 248 | stream.wait_stream(torch.cuda.current_stream()) |
| 249 | with torch.cuda.stream(stream): |
| 250 | run() |
| 251 | torch.cuda.current_stream().wait_stream(stream) |
| 252 | graph = torch.cuda.CUDAGraph() |
| 253 | with torch.cuda.graph(graph): |
| 254 | run() |
| 255 | |
| 256 | for _ in range(3): |
| 257 | raw.normal_() |
| 258 | state_backing.normal_() |
| 259 | history = state.clone() |
| 260 | expected_state = state_backing.clone() |
| 261 | expected_rows = expected_state.as_strided((6, capacity, 1024), (8208, 1024, 1)) |
| 262 | cache_backing.fill_(165) |
| 263 | expected_cache = cache_backing.clone() |
| 264 | latent.fill_(float("nan")) |
| 265 | # Ratio 1 has no ring; ratio 2 stores each chunk's last `capacity` rows. |
| 266 | for t in range(actual if has_ring else 0): |
| 267 | req = req_ids[t].item() |
| 268 | if query_start_loc[req + 1].item() - t > capacity: |
| 269 | continue |
| 270 | slot = state_slots[t].item() |
| 271 | expected_rows[slot // capacity, slot % capacity, :512] = raw[t, :512] |
| 272 | expected_rows[slot // capacity, slot % capacity, 512:] = raw[t, 512:] |
| 273 | |
| 274 | if graph is None: |
| 275 | run() |
| 276 | else: |
| 277 | graph.replay() |
| 278 | |
| 279 | torch.testing.assert_close(state_backing, expected_state, rtol=0, atol=0) |
| 280 | for t in range(num_tokens): |
| 281 | pos = positions[t].item() |
| 282 | if t >= actual or (pos + 1) % compress_ratio: |
| 283 | assert torch.isnan(latent[t]).all() |
| 284 | continue |
| 285 | req = req_ids[t].item() |
| 286 | group_rows = [] |
| 287 | for k in range(compress_ratio - 1, -1, -1): |
| 288 | if t - k >= query_start_loc[req].item(): |
| 289 | kv_score = raw[t - k] |
| 290 | if compress_ratio == 1: |
| 291 | kv_score = torch.cat((kv_score, torch.zeros_like(kv_score))) |
| 292 | group_rows.append(kv_score) |
| 293 | else: |
| 294 | ring = ring_blocks[req].item() |
| 295 | group_rows.append(history[ring, (pos - k) % capacity]) |
| 296 | group = torch.stack(group_rows) |
| 297 | pooled = (group[:, :512] * group[:, 512:].softmax(0)).sum(0) |
| 298 | normed = pooled * torch.rsqrt(pooled.square().mean() + 1e-20) * norm |
| 299 | torch.testing.assert_close(latent[t].float(), normed, rtol=0.004, atol=1e-6) |
| 300 | |
| 301 | slot = cache_slots[t].item() |
| 302 | page, row = divmod(slot, cache_block) |
| 303 | values = expected_cache[page, row * 576 : (row + 1) * 576] |
| 304 | quantized, scales = _ue8m0_reference(latent[t, :448], 64, 448.0) |
| 305 | values[:448] = quantized.view(torch.uint8) |
| 306 | c, s = cos_sin[pos // compress_ratio * compress_ratio].chunk(2) |
| 307 | rope_input = latent[t, 448:].float() |
| 308 | rotated = ( |
| 309 | torch.stack( |
| 310 | ( |
| 311 | rope_input[0::2] * c - rope_input[1::2] * s, |
| 312 | rope_input[1::2] * c + rope_input[0::2] * s, |
| 313 | ), |
| 314 | dim=-1, |
| 315 | ) |
| 316 | .flatten() |
| 317 | .to(torch.bfloat16) |
| 318 | ) |
| 319 | actual_rope = cache_backing[page, row * 576 + 448 : (row + 1) * 576].view( |
| 320 | torch.bfloat16 |
| 321 | ) |
| 322 | torch.testing.assert_close(actual_rope, rotated, rtol=0.008, atol=1e-6) |
| 323 | values[448:] = actual_rope.view(torch.uint8) |
| 324 | scale_offset = cache_block * 576 + row * 8 |
| 325 | expected_cache[page, scale_offset : scale_offset + 7] = ( |
| 326 | scales.log2() + 127 |
| 327 | ).to(torch.uint8) |
| 328 | expected_cache[page, scale_offset + 7] = 0 |
| 329 | # Includes unwritten rows, page padding, quantized values and scales. |
| 330 | torch.testing.assert_close(cache_backing, expected_cache, rtol=0, atol=0) |
| 331 | |
| 332 | |
| 333 | def _rotate_rope_tail( |
| 334 | latent_row: torch.Tensor, cos_sin_row: torch.Tensor |
| 335 | ) -> torch.Tensor: |
| 336 | """GPT-J RoPE of one latent row's last 64 dims, in fp32.""" |
| 337 | row = latent_row.float() |
| 338 | c, s = cos_sin_row.chunk(2) |
| 339 | rotated = row.clone() |
| 340 | even, odd = row[448::2], row[449::2] |
| 341 | rotated[448::2] = even * c - odd * s |
| 342 | rotated[449::2] = odd * c + even * s |
| 343 | return rotated |
| 344 | |
| 345 | |
| 346 | def _mxfp8_record_reference( |
| 347 | latent_row: torch.Tensor, cos_sin_row: torch.Tensor |
| 348 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 349 | """Reference V4.1 record for one token: 512 fp8 bytes + 16 UE8M0 scales. |
| 350 | |
| 351 | Mirrors FlashMLA's ``KVCacheLayout.V41_FP8Sparse`` quantizer: rotate the |
| 352 | RoPE tail first, then scale every 32-dim tile, RoPE tiles included. |
| 353 | """ |
| 354 | rotated = _rotate_rope_tail(latent_row, cos_sin_row) |
| 355 | quantized, scales = _ue8m0_reference(rotated, 32, 448.0) |
| 356 | return quantized.view(torch.uint8), (scales.log2() + 127).to(torch.uint8) |
| 357 | |
| 358 | |
| 359 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 360 | @pytest.mark.parametrize("compress_ratio", [1, 2]) |
| 361 | def test_v41_rope_insert_mxfp8_record(compress_ratio: int): |
| 362 | """The 528-byte V4.1 record quantizes the RoPE dims instead of keeping them |
| 363 | in bf16, so a page is [block_size x 512 fp8][block_size x 16 UE8M0]. |
| 364 | |
| 365 | Byte-exact against the reference quantizer, including the rows and page |
| 366 | padding the kernel must leave alone. |
| 367 | """ |
| 368 | from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import ( |
| 369 | rope_quant_insert, |
| 370 | ) |
| 371 | |
| 372 | torch.manual_seed(7) |
| 373 | device = "cuda" |
| 374 | cache_block = 64 |
| 375 | cache_stride = math.ceil(cache_block * 528 / 512) * 512 |
| 376 | num_tokens = 12 |
| 377 | |
| 378 | positions = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 379 | latent = torch.randn(num_tokens, 512, dtype=torch.bfloat16, device=device) |
| 380 | angles = torch.randn(64, 32, device=device) |
| 381 | cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1) |
| 382 | |
| 383 | cache_backing = torch.full((2, cache_stride), 165, dtype=torch.uint8, device=device) |
| 384 | cache = cache_backing.as_strided((2, cache_block, 528), (cache_stride, 528, 1)) |
| 385 | # Leave one token unmapped so the kernel's negative-slot skip is covered. |
| 386 | slots = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 387 | slots[3] = -1 |
| 388 | |
| 389 | rope_quant_insert(latent, positions, cos_sin, cache, slots, compress_ratio) |
| 390 | |
| 391 | expected = torch.full_like(cache_backing, 165) |
| 392 | for t in range(num_tokens): |
| 393 | slot = slots[t].item() |
| 394 | pos = positions[t].item() |
| 395 | if slot < 0 or (pos + 1) % compress_ratio: |
| 396 | continue |
| 397 | page, row = divmod(slot, cache_block) |
| 398 | values, scales = _mxfp8_record_reference( |
| 399 | latent[t], cos_sin[pos // compress_ratio * compress_ratio] |
| 400 | ) |
| 401 | expected[page, row * 512 : (row + 1) * 512] = values |
| 402 | scale_offset = cache_block * 512 + row * 16 |
| 403 | expected[page, scale_offset : scale_offset + 16] = scales |
| 404 | |
| 405 | torch.testing.assert_close(cache_backing, expected, rtol=0, atol=0) |
| 406 | |
| 407 | |
| 408 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 409 | def test_v41_mxfp8_cache_round_trip(): |
| 410 | """quantize_and_insert -> dequantize_and_gather recovers the V4.1 record. |
| 411 | |
| 412 | The RoPE dims now go through fp8, so they carry quantization error too -- |
| 413 | the check is that every dim survives within one MXFP8 tile step, and that |
| 414 | the bytes on the way match the reference quantizer. |
| 415 | """ |
| 416 | from vllm.models.deepseek_v41.common.ops import ( |
| 417 | dequantize_and_gather_k_cache, |
| 418 | quantize_and_insert_k_cache, |
| 419 | ) |
| 420 | |
| 421 | torch.manual_seed(11) |
| 422 | device = "cuda" |
| 423 | block_size = 64 |
| 424 | num_tokens = 70 |
| 425 | num_blocks = 4 |
| 426 | page_bytes = math.ceil(block_size * 528 / 512) * 512 |
| 427 | |
| 428 | k = torch.randn(num_tokens, 512, dtype=torch.bfloat16, device=device) |
| 429 | cache = torch.zeros(num_blocks, page_bytes, dtype=torch.uint8, device=device) |
| 430 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 431 | quantize_and_insert_k_cache( |
| 432 | k, cache, slot_mapping, block_size=block_size, bytes_per_token=528 |
| 433 | ) |
| 434 | |
| 435 | for t in (0, 1, block_size, num_tokens - 1): |
| 436 | values, scales = _ue8m0_reference(k[t].float(), 32, 448.0) |
| 437 | page, row = divmod(t, block_size) |
| 438 | torch.testing.assert_close( |
| 439 | cache[page, row * 512 : (row + 1) * 512], |
| 440 | values.view(torch.uint8), |
| 441 | rtol=0, |
| 442 | atol=0, |
| 443 | ) |
| 444 | scale_offset = block_size * 512 + row * 16 |
| 445 | torch.testing.assert_close( |
| 446 | cache[page, scale_offset : scale_offset + 16], |
| 447 | (scales.log2() + 127).to(torch.uint8), |
| 448 | rtol=0, |
| 449 | atol=0, |
| 450 | ) |
| 451 | |
| 452 | out = torch.zeros(1, num_tokens, 512, dtype=torch.bfloat16, device=device) |
| 453 | dequantize_and_gather_k_cache( |
| 454 | out, |
| 455 | cache.view(num_blocks, block_size, 528), |
| 456 | seq_lens=torch.tensor([num_tokens], dtype=torch.int32, device=device), |
| 457 | gather_lens=None, |
| 458 | block_table=torch.arange(num_blocks, dtype=torch.int32, device=device).view( |
| 459 | 1, -1 |
| 460 | ), |
| 461 | block_size=block_size, |
| 462 | offset=0, |
| 463 | ) |
| 464 | # Half an e4m3 ULP at the top of the range (16 units) bounds the error. |
| 465 | tile_amax = k.float().abs().view(num_tokens, 16, 32).amax(-1).clamp(min=1e-4) |
| 466 | tolerance = 16.0 * (tile_amax / 448.0).log2().ceil().exp2() |
| 467 | error = (out[0].float() - k.float()).abs() |
| 468 | assert (error <= tolerance.repeat_interleave(32, dim=-1)).all() |
| 469 | |
| 470 | |
| 471 | # e2m1 magnitudes indexed by the low 3 bits of a code; bit 3 is the sign. |
| 472 | _E2M1_MAGNITUDES = [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0] |
| 473 | |
| 474 | |
| 475 | def _decode_nvfp4_row(packed: torch.Tensor, scales: torch.Tensor) -> torch.Tensor: |
| 476 | """Unpack e2m1 rows and apply their e4m3 tile scales.""" |
| 477 | mags = torch.tensor(_E2M1_MAGNITUDES, device=packed.device) |
| 478 | codes = torch.empty( |
| 479 | (*packed.shape[:-1], 512), dtype=torch.uint8, device=packed.device |
| 480 | ) |
| 481 | codes[..., 0::2] = packed & 0xF # even element in the low nibble |
| 482 | codes[..., 1::2] = packed >> 4 |
| 483 | vals = mags[(codes & 7).long()] * torch.where(codes >= 8, -1.0, 1.0) |
| 484 | return ( |
| 485 | vals.reshape(*packed.shape[:-1], 32, 16) |
| 486 | * scales.view(torch.float8_e4m3fn).float().reshape(*packed.shape[:-1], 32, 1) |
| 487 | ).flatten(-2) |
| 488 | |
| 489 | |
| 490 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 491 | @pytest.mark.parametrize("compress_ratio", [1, 2]) |
| 492 | def test_v41_rope_insert_nvfp4_record(compress_ratio: int): |
| 493 | """The 288-byte V4.1 NVFP4 record: 256 B of e2m1 pairs then 32 e4m3 scales. |
| 494 | |
| 495 | Scales are byte-exact against the reference (``amax / 6`` clamped to the |
| 496 | e4m3 range); values are checked as round-to-nearest on the e2m1 grid, whose |
| 497 | coarsest step is 2 between 4 and 6, so half a step is ``1.0 * scale``. |
| 498 | """ |
| 499 | from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import ( |
| 500 | rope_quant_insert, |
| 501 | ) |
| 502 | |
| 503 | torch.manual_seed(13) |
| 504 | device = "cuda" |
| 505 | cache_block = 64 |
| 506 | cache_stride = math.ceil(cache_block * 288 / 512) * 512 |
| 507 | num_tokens = 12 |
| 508 | |
| 509 | positions = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 510 | latent = torch.randn(num_tokens, 512, dtype=torch.bfloat16, device=device) |
| 511 | angles = torch.randn(64, 32, device=device) |
| 512 | cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1) |
| 513 | |
| 514 | cache_backing = torch.full((2, cache_stride), 165, dtype=torch.uint8, device=device) |
| 515 | cache = cache_backing.as_strided((2, cache_block, 288), (cache_stride, 288, 1)) |
| 516 | slots = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 517 | slots[3] = -1 # covers the negative-slot skip |
| 518 | |
| 519 | rope_quant_insert(latent, positions, cos_sin, cache, slots, compress_ratio) |
| 520 | |
| 521 | untouched = torch.full_like(cache_backing, 165) |
| 522 | for t in range(num_tokens): |
| 523 | slot = slots[t].item() |
| 524 | pos = positions[t].item() |
| 525 | if slot < 0 or (pos + 1) % compress_ratio: |
| 526 | continue |
| 527 | page, row = divmod(slot, cache_block) |
| 528 | cs = cos_sin[pos // compress_ratio * compress_ratio] |
| 529 | rotated = _rotate_rope_tail(latent[t], cs) |
| 530 | amax = rotated.view(32, 16).abs().amax(-1) |
| 531 | scale = (amax / 6.0).clamp(2.0**-9, 448.0).to(torch.float8_e4m3fn) |
| 532 | |
| 533 | got_scales = cache_backing[page, cache_block * 256 + row * 32 :][:32] |
| 534 | torch.testing.assert_close(got_scales, scale.view(torch.uint8), rtol=0, atol=0) |
| 535 | decoded = _decode_nvfp4_row( |
| 536 | cache_backing[page, row * 256 : (row + 1) * 256], got_scales |
| 537 | ) |
| 538 | step = scale.float().repeat_interleave(16) |
| 539 | assert ((decoded - rotated).abs() <= step).all() |
| 540 | untouched[page, row * 256 : (row + 1) * 256] = cache_backing[ |
| 541 | page, row * 256 : (row + 1) * 256 |
| 542 | ] |
| 543 | untouched[page, cache_block * 256 + row * 32 :][:32] = got_scales |
| 544 | # Rows the kernel must not have touched, page padding included. |
| 545 | torch.testing.assert_close(cache_backing, untouched, rtol=0, atol=0) |
| 546 | |
| 547 | |
| 548 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 549 | @pytest.mark.parametrize("num_reqs", [1, 8, 32]) |
| 550 | @pytest.mark.parametrize("num_tokens", [70, 128, 129, 2051]) |
| 551 | @pytest.mark.parametrize("partial", [False, True]) |
| 552 | def test_v41_nvfp4_gather_matches_insert(num_reqs, num_tokens, partial): |
| 553 | """The NVFP4 gather dequantizes exactly what the insert kernel wrote.""" |
| 554 | from vllm.models.deepseek_v41.common.ops import dequantize_and_gather_k_cache |
| 555 | from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import ( |
| 556 | rope_quant_insert, |
| 557 | ) |
| 558 | |
| 559 | torch.manual_seed(17) |
| 560 | device = "cuda" |
| 561 | block_size = 64 |
| 562 | num_blocks = math.ceil(num_tokens / block_size) |
| 563 | page_bytes = math.ceil(block_size * 288 / 512) * 512 |
| 564 | |
| 565 | positions = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 566 | latent = torch.randn(num_tokens, 512, dtype=torch.bfloat16, device=device) |
| 567 | angles = torch.randn(num_tokens, 32, device=device) |
| 568 | cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1) |
| 569 | backing = torch.zeros(num_blocks, page_bytes, dtype=torch.uint8, device=device) |
| 570 | cache = backing.as_strided((num_blocks, block_size, 288), (page_bytes, 288, 1)) |
| 571 | slots = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 572 | rope_quant_insert(latent, positions, cos_sin, cache, slots, 1) |
| 573 | |
| 574 | offset = 7 if partial else 0 |
| 575 | seq_lens = num_tokens - torch.arange(num_reqs, dtype=torch.int32, device=device) |
| 576 | gather_lens = seq_lens - 3 if partial else None |
| 577 | storage = torch.full( |
| 578 | (num_reqs, num_tokens + offset + 16, 512), |
| 579 | -123.0, |
| 580 | dtype=torch.bfloat16, |
| 581 | device=device, |
| 582 | ) |
| 583 | out = storage[:, : num_tokens + offset] |
| 584 | block_table = torch.arange(num_blocks, dtype=torch.int32, device=device).repeat( |
| 585 | num_reqs, 1 |
| 586 | ) |
| 587 | |
| 588 | def gather(): |
| 589 | dequantize_and_gather_k_cache( |
| 590 | out, |
| 591 | cache, |
| 592 | seq_lens=seq_lens, |
| 593 | gather_lens=gather_lens, |
| 594 | block_table=block_table, |
| 595 | block_size=block_size, |
| 596 | offset=offset, |
| 597 | ) |
| 598 | |
| 599 | expected = _decode_nvfp4_row( |
| 600 | backing[:, : block_size * 256].reshape(-1, 256), |
| 601 | backing[:, block_size * 256 : block_size * 288].reshape(-1, 32), |
| 602 | ) |
| 603 | |
| 604 | def check(shorter_by): |
| 605 | assert (storage[:, num_tokens + offset :] == -123).all() |
| 606 | for req in range(num_reqs): |
| 607 | start = 3 if partial else 0 |
| 608 | length = num_tokens - req - start - shorter_by |
| 609 | assert (out[req, :offset] == -123).all() |
| 610 | assert (out[req, offset + length :] == -123).all() |
| 611 | torch.testing.assert_close( |
| 612 | out[req, offset : offset + length].float(), |
| 613 | expected[start : start + length], |
| 614 | rtol=0, |
| 615 | atol=0, |
| 616 | ) |
| 617 | |
| 618 | gather() |
| 619 | check(0) |
| 620 | graph = torch.cuda.CUDAGraph() |
| 621 | with torch.cuda.graph(graph): |
| 622 | gather() |
| 623 | storage.fill_(-123) |
| 624 | seq_lens.sub_(1) |
| 625 | if gather_lens is not None: |
| 626 | gather_lens.sub_(1) |
| 627 | graph.replay() |
| 628 | check(1) |
| 629 | |
| 630 | |
| 631 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 632 | @pytest.mark.parametrize("compress_ratio", [1, 2]) |
| 633 | @pytest.mark.parametrize("store_fp8", [False, True]) |
| 634 | def test_v41_rope_insert_plain_row(compress_ratio: int, store_fp8: bool): |
| 635 | """The FlashInfer plain-row compressed cache gets [448 NoPE | 64 RoPE] rows. |
| 636 | |
| 637 | bf16 caches store the latent verbatim with the RoPE tail rotated; per-tensor |
| 638 | fp8 caches scale the bf16-rounded row by 1/fp8_scale and clamp to e4m3. |
| 639 | Non-boundary positions, negative slots and untouched rows must keep their |
| 640 | prior contents. |
| 641 | """ |
| 642 | from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import ( |
| 643 | rope_quant_insert, |
| 644 | ) |
| 645 | |
| 646 | torch.manual_seed(3) |
| 647 | device = "cuda" |
| 648 | num_tokens = 9 |
| 649 | cache_block = 4 |
| 650 | positions = torch.arange(5, 5 + num_tokens, dtype=torch.int64, device=device) |
| 651 | cache_slots = torch.randperm(3 * cache_block, device=device)[:num_tokens] |
| 652 | cache_slots[2] = -1 |
| 653 | latent = torch.randn(num_tokens, 512, device=device).to(torch.bfloat16) |
| 654 | angles = torch.randn(32, 32, device=device) |
| 655 | cos_sin = torch.cat((angles.cos(), angles.sin()), dim=-1) |
| 656 | dtype = torch.float8_e4m3fn if store_fp8 else torch.bfloat16 |
| 657 | # Pad the page stride so the kernel must honour strides, not shape. |
| 658 | cache_backing = torch.full((3, (cache_block + 1) * 512), 3.0, device=device) |
| 659 | cache_backing = cache_backing.to(dtype) |
| 660 | cache = cache_backing.as_strided( |
| 661 | (3, cache_block, 512), ((cache_block + 1) * 512, 512, 1) |
| 662 | ) |
| 663 | fp8_scale = torch.tensor([0.5], dtype=torch.float32, device=device) |
| 664 | expected = cache_backing.clone().float() |
| 665 | |
| 666 | rope_quant_insert( |
| 667 | latent, |
| 668 | positions, |
| 669 | cos_sin, |
| 670 | cache, |
| 671 | cache_slots, |
| 672 | compress_ratio, |
| 673 | fp8_scale=fp8_scale if store_fp8 else None, |
| 674 | ) |
| 675 | |
| 676 | for t in range(num_tokens): |
| 677 | pos = positions[t].item() |
| 678 | slot = cache_slots[t].item() |
| 679 | if slot < 0 or (pos + 1) % compress_ratio: |
| 680 | continue |
| 681 | c, s = cos_sin[pos // compress_ratio * compress_ratio].chunk(2) |
| 682 | rope_input = latent[t, 448:].float() |
| 683 | rotated = torch.stack( |
| 684 | ( |
| 685 | rope_input[0::2] * c - rope_input[1::2] * s, |
| 686 | rope_input[1::2] * c + rope_input[0::2] * s, |
| 687 | ), |
| 688 | dim=-1, |
| 689 | ).flatten() |
| 690 | row = torch.cat((latent[t, :448], rotated.to(torch.bfloat16))).float() |
| 691 | if store_fp8: |
| 692 | row = (row * (1.0 / fp8_scale)).clamp(-448.0, 448.0) |
| 693 | row = row.to(torch.float8_e4m3fn).float() |
| 694 | page, idx = divmod(slot, cache_block) |
| 695 | expected[page, idx * 512 : (idx + 1) * 512] = row |
| 696 | |
| 697 | actual = cache_backing.float() |
| 698 | if store_fp8: |
| 699 | # One e4m3 ulp of slack for FMA-vs-separate rounding in the RoPE tail. |
| 700 | torch.testing.assert_close(actual, expected, rtol=0.13, atol=1e-2) |
| 701 | else: |
| 702 | nope = torch.arange(512, device=device) < 448 |
| 703 | nope = nope.repeat(cache_block + 1) |
| 704 | torch.testing.assert_close(actual[:, nope], expected[:, nope], rtol=0, atol=0) |
| 705 | torch.testing.assert_close( |
| 706 | actual[:, ~nope], expected[:, ~nope], rtol=0.008, atol=1e-6 |
| 707 | ) |
| 708 | |
| 709 | |
| 710 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="Triton kernel") |
| 711 | def test_v41_compressor_metadata_maps_tokens_to_their_ring(): |
| 712 | """The ring group's generic slot mapping is disabled (all PAD), so the |
| 713 | builder must map every real token to ``ring_block * capacity + pos % |
| 714 | capacity`` and keep padding tokens at PAD.""" |
| 715 | from unittest.mock import MagicMock |
| 716 | |
| 717 | from vllm.models.deepseek_v41.compressor import CompressorMetadataBuilder |
| 718 | from vllm.v1.attention.backend import CommonAttentionMetadata |
| 719 | from vllm.v1.kv_cache_interface import CircularBufferSpec |
| 720 | |
| 721 | capacity = 8 |
| 722 | vllm_config = MagicMock() |
| 723 | vllm_config.scheduler_config.max_num_batched_tokens = 16 |
| 724 | spec = CircularBufferSpec( |
| 725 | block_size=capacity, |
| 726 | num_kv_heads=1, |
| 727 | head_size=1024, |
| 728 | head_size_v=0, |
| 729 | dtype=torch.float32, |
| 730 | ) |
| 731 | device = torch.device("cuda") |
| 732 | builder = CompressorMetadataBuilder(spec, ["state"], vllm_config, device) |
| 733 | |
| 734 | # Two requests: 3 tokens at positions 13..15 on ring block 5, then 2 |
| 735 | # tokens at positions 7..8 on ring block 2; three padding tokens. |
| 736 | query_start_loc = torch.tensor([0, 3, 5], dtype=torch.int32, device=device) |
| 737 | positions = torch.tensor([13, 14, 15, 7, 8, 0, 0, 0], device=device) |
| 738 | block_table = torch.tensor([[5], [2]], dtype=torch.int32, device=device) |
| 739 | common = CommonAttentionMetadata( |
| 740 | query_start_loc=query_start_loc, |
| 741 | query_start_loc_cpu=query_start_loc.cpu(), |
| 742 | seq_lens=torch.tensor([16, 9], dtype=torch.int32, device=device), |
| 743 | num_reqs=2, |
| 744 | num_actual_tokens=5, |
| 745 | max_query_len=3, |
| 746 | max_seq_len=16, |
| 747 | block_table_tensor=block_table, |
| 748 | slot_mapping=torch.full((8,), -1, dtype=torch.int64, device=device), |
| 749 | positions=positions, |
| 750 | ) |
| 751 | metadata = builder.build(0, common) |
| 752 | |
| 753 | expected = [5 * 8 + 5, 5 * 8 + 6, 5 * 8 + 7, 2 * 8 + 7, 2 * 8 + 0, -1, -1, -1] |
| 754 | assert metadata.slot_mapping.tolist() == expected |
| 755 | assert metadata.query_start_loc is query_start_loc |
| 756 | assert metadata.token_to_req_indices.tolist() == [0, 0, 0, 1, 1] |
| 757 | |
| 758 | |
| 759 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA stream coverage") |
| 760 | @pytest.mark.parametrize( |
| 761 | "use_aux,use_graph", [(False, False), (True, False), (True, True)] |
| 762 | ) |
| 763 | @torch.inference_mode() |
| 764 | def test_v41_attention_joins_cache_writes_before_consumption(use_aux, use_graph): |
| 765 | """Both reused-event joins must publish this forward's states and cache rows.""" |
| 766 | from vllm.forward_context import ForwardContext, override_forward_context |
| 767 | from vllm.models.deepseek_v41.attention import ( |
| 768 | DeepseekV4Attention, |
| 769 | DeepseekV4Indexer, |
| 770 | ) |
| 771 | from vllm.models.deepseek_v41.compressor import DeepseekCompressor |
| 772 | |
| 773 | torch.manual_seed(43) |
| 774 | raw = torch.randn(19, 1024, device="cuda") |
| 775 | q = torch.zeros(19, 512, dtype=torch.bfloat16, device="cuda") |
| 776 | positions = torch.arange(7, 26, device="cuda") |
| 777 | state_slots = positions.clone() |
| 778 | state_slots[-2:] = -1 |
| 779 | cache_slots = torch.where(state_slots >= 0, positions // 2, -1) |
| 780 | state = torch.empty(4, 8, 1024, device="cuda") |
| 781 | main = torch.empty(1, 128, 584, dtype=torch.uint8, device="cuda") |
| 782 | index = torch.empty(1, 128, 132, dtype=torch.uint8, device="cuda") |
| 783 | caches = (state, main, index) |
| 784 | observed = tuple(torch.empty_like(cache) for cache in caches) |
| 785 | rotary = SimpleNamespace(cos_sin_cache=torch.randn(32, 64, device="cuda")) |
| 786 | metadata = { |
| 787 | "state": SimpleNamespace( |
| 788 | slot_mapping=state_slots, |
| 789 | query_start_loc=torch.tensor([0, 19], dtype=torch.int32, device="cuda"), |
| 790 | token_to_req_indices=torch.zeros(19, dtype=torch.int32, device="cuda"), |
| 791 | ), |
| 792 | "main": SimpleNamespace(slot_mapping=cache_slots), |
| 793 | "index": SimpleNamespace(slot_mapping=cache_slots), |
| 794 | } |
| 795 | context = ForwardContext({}, metadata, {}) |
| 796 | compressor = DeepseekCompressor.__new__(DeepseekCompressor) |
| 797 | torch.nn.Module.__init__(compressor) |
| 798 | compressor.head_dim, compressor.rope_head_dim, compressor.compress_ratio = ( |
| 799 | 512, |
| 800 | 64, |
| 801 | 2, |
| 802 | ) |
| 803 | compressor.rms_norm_eps = 1e-20 |
| 804 | compressor.norm = SimpleNamespace( |
| 805 | weight=torch.ones(512, dtype=torch.bfloat16, device="cuda") |
| 806 | ) |
| 807 | compressor.state_cache = SimpleNamespace(prefix="state", kv_cache=state) |
| 808 | compressor.k_cache_prefix = "main" |
| 809 | compressor._static_forward_context = {"main": SimpleNamespace(kv_cache=main)} |
| 810 | indexer_weight = torch.randn(128, 512, dtype=torch.bfloat16, device="cuda") |
| 811 | indexer = SimpleNamespace( |
| 812 | owns_k=True, |
| 813 | wk=lambda latent: (torch.nn.functional.linear(latent, indexer_weight), None), |
| 814 | k_norm=SimpleNamespace( |
| 815 | weight=torch.ones(128, dtype=torch.bfloat16, device="cuda"), |
| 816 | variance_epsilon=1e-20, |
| 817 | ), |
| 818 | k_cache=SimpleNamespace(prefix="index", kv_cache=index), |
| 819 | compress_ratio=2, |
| 820 | use_fp4_kv=False, |
| 821 | ) |
| 822 | |
| 823 | def prepare_indexer(qr, latent, weights, positions, rotary, qr_scale): |
| 824 | DeepseekV4Indexer._produce_k(indexer, latent, positions, rotary) |
| 825 | return None, None, None |
| 826 | |
| 827 | def observe(*args): |
| 828 | for output, cache in zip(observed, caches): |
| 829 | output.copy_(cache) |
| 830 | |
| 831 | auxiliary = [torch.cuda.Stream()] if use_aux else None |
| 832 | attention = SimpleNamespace( |
| 833 | compressor=compressor, |
| 834 | indexer=prepare_indexer, |
| 835 | aux_stream_list=auxiliary, |
| 836 | ln_events=[torch.cuda.Event(), torch.cuda.Event()], |
| 837 | rotary_emb=rotary, |
| 838 | indexer_rotary_emb=rotary, |
| 839 | n_local_heads=1, |
| 840 | head_dim=512, |
| 841 | _wq_b_proj=lambda qr, scale: qr.clone(), |
| 842 | _fused_qnorm_rope_kv_insert=lambda q, kv, pos, meta: q, |
| 843 | _sparse_indexer_and_attn=observe, |
| 844 | ) |
| 845 | |
| 846 | def run(): |
| 847 | with override_forward_context(context): |
| 848 | DeepseekV4Attention._prepare_and_attn( |
| 849 | attention, q, q, q, None, raw, q, positions, q |
| 850 | ) |
| 851 | |
| 852 | def reset(): |
| 853 | state.zero_() |
| 854 | main.fill_(165) |
| 855 | index.fill_(165) |
| 856 | |
| 857 | reset() |
| 858 | run() |
| 859 | graph = None |
| 860 | if use_graph: |
| 861 | stream = torch.cuda.Stream() |
| 862 | stream.wait_stream(torch.cuda.current_stream()) |
| 863 | with torch.cuda.stream(stream): |
| 864 | run() |
| 865 | torch.cuda.current_stream().wait_stream(stream) |
| 866 | graph = torch.cuda.CUDAGraph() |
| 867 | with torch.cuda.graph(graph): |
| 868 | run() |
| 869 | |
| 870 | previous = None |
| 871 | for _ in range(3): |
| 872 | raw.normal_() |
| 873 | reset() |
| 874 | attention.aux_stream_list = None |
| 875 | run() |
| 876 | expected = tuple(output.clone() for output in observed) |
| 877 | reset() |
| 878 | attention.aux_stream_list = auxiliary |
| 879 | if graph is None: |
| 880 | run() |
| 881 | else: |
| 882 | graph.replay() |
| 883 | for actual, reference in zip(observed, expected): |
| 884 | torch.testing.assert_close(actual, reference, rtol=0, atol=0) |
| 885 | if previous is not None: |
| 886 | assert all(not torch.equal(a, b) for a, b in zip(expected, previous)) |
| 887 | previous = expected |
| 888 | |
| 889 | |
| 890 | def _on_gfx950() -> bool: |
| 891 | if not current_platform.is_rocm(): |
| 892 | return False |
| 893 | try: |
| 894 | from vllm.platforms.rocm import _ON_GFX950 |
| 895 | |
| 896 | return _ON_GFX950 |
| 897 | except Exception: |
| 898 | return False |
| 899 | |
| 900 | |
| 901 | @pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm-only dispatch") |
| 902 | def test_cp_gather_despecialized_kernel_is_gfx950_only(monkeypatch): |
| 903 | from vllm.v1.attention.ops import rocm_aiter_mla_sparse as mod |
| 904 | |
| 905 | class FakeKernel: |
| 906 | def __init__(self): |
| 907 | self.calls = [] |
| 908 | |
| 909 | def __getitem__(self, grid): |
| 910 | def launch(*args): |
| 911 | self.calls.append((grid, args)) |
| 912 | |
| 913 | return launch |
| 914 | |
| 915 | legacy_kernel = FakeKernel() |
| 916 | gfx950_kernel = FakeKernel() |
| 917 | monkeypatch.setattr(mod, "_cp_gather_indexer_quant_cache_kernel", legacy_kernel) |
| 918 | monkeypatch.setattr( |
| 919 | mod, |
| 920 | "_cp_gather_indexer_quant_cache_gfx950_kernel", |
| 921 | gfx950_kernel, |
| 922 | ) |
| 923 | |
| 924 | k_cache = torch.zeros((4, 1, 132), dtype=torch.uint8) |
| 925 | k_fp8 = torch.empty((5, 128), dtype=current_platform.fp8_dtype()) |
| 926 | k_scale = torch.empty((5, 4), dtype=torch.uint8) |
| 927 | block_table = torch.zeros((2, 7), dtype=torch.int32) |
| 928 | cu_seqlen = torch.tensor([0, 2, 5], dtype=torch.int32) |
| 929 | token_to_seq = torch.tensor([0, 0, 1, 1, 1], dtype=torch.int32) |
| 930 | args = (k_cache, k_fp8, k_scale, block_table, cu_seqlen, token_to_seq) |
| 931 | |
| 932 | monkeypatch.setattr(mod, "_ON_GFX950", True) |
| 933 | mod.cp_gather_indexer_k_quant_cache_triton(*args) |
| 934 | assert len(gfx950_kernel.calls) == 1 |
| 935 | assert not legacy_kernel.calls |
| 936 | gfx950_grid, gfx950_args = gfx950_kernel.calls[0] |
| 937 | assert gfx950_grid == (5,) |
| 938 | assert len(gfx950_args) == 18 |
| 939 | assert gfx950_args[-3:] == (2, 7, 4) |
| 940 | |
| 941 | monkeypatch.setattr(mod, "_ON_GFX950", False) |
| 942 | mod.cp_gather_indexer_k_quant_cache_triton(*args) |
| 943 | assert len(legacy_kernel.calls) == 1 |
| 944 | legacy_grid, legacy_args = legacy_kernel.calls[0] |
| 945 | assert legacy_grid == (5,) |
| 946 | assert len(legacy_args) == 19 |
| 947 | assert legacy_args[-4:] == (5, 2, 7, 4) |
| 948 | |
| 949 | |
| 950 | @pytest.mark.parametrize( |
| 951 | ("window_size", "num_speculative_tokens", "expected"), |
| 952 | [(128, 5, 192), (512, 5, 576), (1024, 0, 1024)], |
| 953 | ) |
| 954 | def test_get_dspark_swa_index_width( |
| 955 | window_size: int, num_speculative_tokens: int, expected: int |
| 956 | ): |
| 957 | assert get_dspark_swa_index_width(window_size, num_speculative_tokens) == expected |
| 958 | |
| 959 | |
| 960 | def _ue8m0_reference(x: torch.Tensor, block_size: int, fp8_max: float): |
| 961 | """PyTorch reference for UE8M0 FP8 quantization (per-block, power-of-2 scale). |
| 962 | |
| 963 | Returns (x_fp8, scales) where x_fp8 is float8_e4m3fn and scales are float32. |
| 964 | """ |
| 965 | assert x.dim() == 1 |
| 966 | n = x.numel() |
| 967 | n_blocks = math.ceil(n / block_size) |
| 968 | x_fp8 = torch.zeros(n, dtype=torch.float8_e4m3fn, device=x.device) |
| 969 | scales = torch.zeros(n_blocks, dtype=torch.float32, device=x.device) |
| 970 | |
| 971 | for i in range(n_blocks): |
| 972 | start = i * block_size |
| 973 | end = min(start + block_size, n) |
| 974 | block = x[start:end].float() |
| 975 | amax = block.abs().max().clamp(min=1e-4) |
| 976 | raw_scale = amax / fp8_max |
| 977 | exponent = math.ceil(math.log2(raw_scale.item())) |
| 978 | scale = 2.0**exponent |
| 979 | scales[i] = scale |
| 980 | quantized = (block / scale).clamp(-fp8_max, fp8_max) |
| 981 | x_fp8[start:end] = quantized.to(torch.float8_e4m3fn) |
| 982 | |
| 983 | return x_fp8, scales |
| 984 | |
| 985 | |
| 986 | def _decode_dsv4_cache_row( |
| 987 | cache: torch.Tensor, block_size: int, scrub_nan: bool |
| 988 | ) -> torch.Tensor: |
| 989 | flat = cache.flatten() |
| 990 | nope = flat[:448].view(torch.float8_e4m3fn).to(torch.bfloat16) |
| 991 | encoded = flat[block_size * 576 : block_size * 576 + 7] |
| 992 | scales = torch.exp2(encoded.to(torch.float32) - 127.0).to(torch.bfloat16) |
| 993 | nope = nope * scales.repeat_interleave(64) |
| 994 | rope = flat[448:576].view(torch.bfloat16) |
| 995 | decoded = torch.cat((nope, rope)) |
| 996 | if scrub_nan: |
| 997 | decoded = torch.where(decoded == decoded, decoded, 0.0) |
| 998 | return decoded |
| 999 | |
| 1000 | |
| 1001 | def _assert_nan_free_cache_matches_legacy_scrub( |
| 1002 | cache: torch.Tensor, block_size: int |
| 1003 | ) -> None: |
| 1004 | flat = cache.flatten() |
| 1005 | scale_base = block_size * 576 |
| 1006 | scale_codes = flat[scale_base : scale_base + 8] |
| 1007 | assert scale_codes[0].item() == 254 |
| 1008 | assert scale_codes[1].item() == 247 |
| 1009 | assert scale_codes[:7].max().item() <= 254 |
| 1010 | nope_bytes = flat[:448] |
| 1011 | assert not ((nope_bytes == 0x7F) | (nope_bytes == 0xFF)).any() |
| 1012 | |
| 1013 | rope = flat[448:576].view(torch.bfloat16) |
| 1014 | assert not torch.isnan(rope).any() |
| 1015 | assert torch.isposinf(rope[0]) |
| 1016 | assert torch.equal(rope[1:4], torch.zeros_like(rope[1:4])) |
| 1017 | |
| 1018 | legacy_cache = cache.clone() |
| 1019 | legacy_flat = legacy_cache.flatten() |
| 1020 | legacy_flat[scale_base] = 255 |
| 1021 | legacy_rope = legacy_flat[448:576].view(torch.bfloat16) |
| 1022 | legacy_rope[1:4] = float("nan") |
| 1023 | canonical = _decode_dsv4_cache_row(cache, block_size, scrub_nan=False) |
| 1024 | legacy = _decode_dsv4_cache_row(legacy_cache, block_size, scrub_nan=True) |
| 1025 | torch.testing.assert_close(canonical, legacy, rtol=0, atol=0) |
| 1026 | assert torch.isinf(canonical[0]) |
| 1027 | assert torch.isposinf(canonical[64]) |
| 1028 | |
| 1029 | |
| 1030 | @pytest.mark.skipif( |
| 1031 | not _on_gfx950(), |
| 1032 | reason="NaN-free fp8_ds_mla compressed-cache contract is gfx950-only", |
| 1033 | ) |
| 1034 | @pytest.mark.parametrize("writer", ["single_pass", "two_stage_finalizer"]) |
| 1035 | def test_gfx950_compressed_cache_canonicalizes_nonfinite(writer: str) -> None: |
| 1036 | head_dim = 512 |
| 1037 | rope_dim = 64 |
| 1038 | block_size = 4 |
| 1039 | device = "cuda" |
| 1040 | |
| 1041 | positions = torch.zeros(1, dtype=torch.int64, device=device) |
| 1042 | slot_mapping = torch.zeros(1, dtype=torch.int64, device=device) |
| 1043 | rms_weight = torch.ones(head_dim, dtype=torch.bfloat16, device=device) |
| 1044 | rms_weight[0] = float("inf") |
| 1045 | rms_weight[64] = torch.finfo(torch.bfloat16).max |
| 1046 | rms_weight[448] = float("inf") |
| 1047 | rms_weight[450] = float("nan") |
| 1048 | cos_sin_cache = torch.zeros(1, rope_dim, dtype=torch.float32, device=device) |
| 1049 | cos_sin_cache[:, : rope_dim // 2] = 1.0 |
| 1050 | cache = torch.zeros(1, block_size, 584, dtype=torch.uint8, device=device) |
| 1051 | |
| 1052 | state_cache = torch.zeros(1, 1, 2 * head_dim, dtype=torch.float32, device=device) |
| 1053 | state_cache[..., :head_dim] = 1.0 |
| 1054 | token_to_req = torch.zeros(1, dtype=torch.int32, device=device) |
| 1055 | block_table = torch.zeros(1, 1, dtype=torch.int32, device=device) |
| 1056 | |
| 1057 | if writer == "single_pass": |
| 1058 | compress_norm_rope_store_triton( |
| 1059 | state_cache=state_cache, |
| 1060 | num_actual=1, |
| 1061 | token_to_req_indices=token_to_req, |
| 1062 | positions=positions, |
| 1063 | slot_mapping=slot_mapping, |
| 1064 | block_table=block_table, |
| 1065 | block_size=1, |
| 1066 | state_width=head_dim, |
| 1067 | cos_sin_cache=cos_sin_cache, |
| 1068 | kv_cache=cache, |
| 1069 | k_cache_metadata=SimpleNamespace(slot_mapping=slot_mapping), |
| 1070 | pdl_kwargs={}, |
| 1071 | head_dim=head_dim, |
| 1072 | rope_head_dim=rope_dim, |
| 1073 | compress_ratio=1, |
| 1074 | overlap=False, |
| 1075 | use_fp4_cache=False, |
| 1076 | rms_norm_weight=rms_weight, |
| 1077 | rms_norm_eps=1e-6, |
| 1078 | quant_block=64, |
| 1079 | token_stride=576, |
| 1080 | scale_dim=8, |
| 1081 | ) |
| 1082 | else: |
| 1083 | _launch_two_stage_sparse_attn_compressor( |
| 1084 | state_cache, |
| 1085 | token_to_req, |
| 1086 | positions, |
| 1087 | slot_mapping, |
| 1088 | block_table, |
| 1089 | 1, |
| 1090 | head_dim, |
| 1091 | 1, |
| 1092 | cos_sin_cache, |
| 1093 | cache, |
| 1094 | slot_mapping, |
| 1095 | rms_weight, |
| 1096 | 1e-6, |
| 1097 | 64, |
| 1098 | 576, |
| 1099 | 8, |
| 1100 | head_dim, |
| 1101 | rope_dim, |
| 1102 | 1, |
| 1103 | torch.empty(1, head_dim, dtype=torch.float32, device=device), |
| 1104 | ) |
| 1105 | |
| 1106 | _assert_nan_free_cache_matches_legacy_scrub(cache, block_size) |
| 1107 | |
| 1108 | |
| 1109 | @pytest.mark.parametrize( |
| 1110 | ("starts", "query_start_loc", "expected"), |
| 1111 | [ |
| 1112 | ([0], [0, 127], False), |
| 1113 | ([0], [0, 128], True), |
| 1114 | ([127], [0, 1], True), |
| 1115 | ([128], [0, 127], False), |
| 1116 | ([1, 255], [0, 1, 2], True), |
| 1117 | (None, [0, 1], None), |
| 1118 | ], |
| 1119 | ) |
| 1120 | def test_get_c128_boundary(starts, query_start_loc, expected): |
| 1121 | query_start_loc_tensor = torch.tensor(query_start_loc) |
| 1122 | query_lens = query_start_loc_tensor[1:] - query_start_loc_tensor[:-1] |
| 1123 | metadata = SimpleNamespace( |
| 1124 | seq_lens_cpu_upper_bound=( |
| 1125 | None if starts is None else torch.tensor(starts) + query_lens |
| 1126 | ), |
| 1127 | query_start_loc_cpu=query_start_loc_tensor, |
| 1128 | ) |
| 1129 | assert _get_c128_boundary(metadata) is expected |
| 1130 | |
| 1131 | |
| 1132 | # ── Test A: DeepseekV4 Attention path ────────────────────────────────────────────── |
| 1133 | |
| 1134 | |
| 1135 | @pytest.mark.parametrize("num_tokens", [1, 4, 8, 17]) |
| 1136 | @pytest.mark.parametrize("block_size", [16, 64]) |
| 1137 | def test_deepseek_v4_attention_quant_cache_roundtrip(num_tokens: int, block_size: int): |
| 1138 | """compressed_kv → quantize_and_insert_k_cache → dequantize_and_gather_k_cache |
| 1139 | → compare against original.""" |
| 1140 | HEAD_DIM = 512 |
| 1141 | NOPE_DIM = 448 |
| 1142 | HEAD_BYTES = 584 # 448 fp8 + 128 bf16 + 8 uint8 scale |
| 1143 | FP8_MAX = 448.0 |
| 1144 | QUANT_BLOCK = 64 |
| 1145 | |
| 1146 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 1147 | device = "cuda" |
| 1148 | |
| 1149 | # Random compressed_kv (simulates compressor output) |
| 1150 | compressed_kv = torch.randn( |
| 1151 | num_tokens, HEAD_DIM, dtype=torch.bfloat16, device=device |
| 1152 | ) |
| 1153 | |
| 1154 | # ── Quant + insert ────────────────────────────────────────────────── |
| 1155 | k_cache = torch.zeros( |
| 1156 | num_blocks, block_size, HEAD_BYTES, dtype=torch.uint8, device=device |
| 1157 | ) |
| 1158 | k_cache_2d = k_cache.view(num_blocks, -1) |
| 1159 | |
| 1160 | # Sequential slot mapping: token i → slot i |
| 1161 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 1162 | |
| 1163 | quantize_and_insert_k_cache( |
| 1164 | compressed_kv, k_cache_2d, slot_mapping, block_size=block_size |
| 1165 | ) |
| 1166 | |
| 1167 | # ── Gather + dequant ──────────────────────────────────────────────── |
| 1168 | num_reqs = 1 |
| 1169 | max_blocks_per_seq = num_blocks |
| 1170 | out = torch.zeros( |
| 1171 | num_reqs, num_tokens, HEAD_DIM, dtype=torch.bfloat16, device=device |
| 1172 | ) |
| 1173 | seq_lens = torch.tensor([num_tokens], dtype=torch.int32, device=device) |
| 1174 | # block_table: request 0 uses physical blocks 0, 1, ... |
| 1175 | block_table = torch.arange( |
| 1176 | max_blocks_per_seq, dtype=torch.int32, device=device |
| 1177 | ).unsqueeze(0) |
| 1178 | |
| 1179 | dequantize_and_gather_k_cache( |
| 1180 | out, k_cache, seq_lens, None, block_table, block_size, offset=0 |
| 1181 | ) |
| 1182 | |
| 1183 | recovered = out[0, :num_tokens] |
| 1184 | |
| 1185 | # ── NoPE portion (first 448): FP8 quantized, expect UE8M0 error ── |
| 1186 | nope_orig = compressed_kv[:, :NOPE_DIM].float() |
| 1187 | nope_recv = recovered[:, :NOPE_DIM].float() |
| 1188 | nope_diff = (nope_recv - nope_orig).abs() |
| 1189 | |
| 1190 | # Per-token check: FP8 e4m3 (3-bit mantissa) worst-case error is |
| 1191 | # half-ULP at the largest representable value. At y ≈ 448 (max), |
| 1192 | # ULP = 2^(8-3) = 32, so error ≤ 16 * scale. |
| 1193 | for t in range(num_tokens): |
| 1194 | _, scales = _ue8m0_reference( |
| 1195 | compressed_kv[t, :NOPE_DIM].float(), QUANT_BLOCK, FP8_MAX |
| 1196 | ) |
| 1197 | max_allowed = 16.0 * scales.max().item() |
| 1198 | token_diff = nope_diff[t].max().item() |
| 1199 | assert token_diff <= max_allowed, ( |
| 1200 | f"Token {t} nope diff {token_diff} exceeds max_allowed " |
| 1201 | f"{max_allowed} (scale={scales.max().item()})" |
| 1202 | ) |
| 1203 | |
| 1204 | # ── RoPE portion (last 64): stored as bf16, should be exact ───── |
| 1205 | rope_diff = (recovered[:, NOPE_DIM:] - compressed_kv[:, NOPE_DIM:]).abs() |
| 1206 | assert rope_diff.max().item() == 0.0, ( |
| 1207 | f"RoPE portion should be exact but got max diff {rope_diff.max().item()}" |
| 1208 | ) |
| 1209 | |
| 1210 | |
| 1211 | # ── Test B: Fused dequant+gather K cache ──────────────────────────────────── |
| 1212 | |
| 1213 | |
| 1214 | def _dequantize_and_gather_k_cache_reference( |
| 1215 | out: torch.Tensor, |
| 1216 | k_cache: torch.Tensor, |
| 1217 | seq_lens: torch.Tensor, |
| 1218 | gather_lens: torch.Tensor | None, |
| 1219 | block_table: torch.Tensor, |
| 1220 | block_size: int, |
| 1221 | offset: int, |
| 1222 | ) -> None: |
| 1223 | fp8_dim = 448 |
| 1224 | bf16_dim = 64 |
| 1225 | scale_dim = 8 |
| 1226 | quant_block = 64 |
| 1227 | token_data_size = fp8_dim + bf16_dim * 2 |
| 1228 | |
| 1229 | for req_id in range(seq_lens.shape[0]): |
| 1230 | seq_len = seq_lens[req_id].item() |
| 1231 | gather_len = gather_lens[req_id].item() if gather_lens is not None else seq_len |
| 1232 | start_pos = seq_len - gather_len |
| 1233 | |
| 1234 | for i in range(gather_len): |
| 1235 | pos = start_pos + i |
| 1236 | pos_in_block = pos % block_size |
| 1237 | block_idx = block_table[req_id, pos // block_size].item() |
| 1238 | cache_block = k_cache[block_idx].view(-1) |
| 1239 | |
| 1240 | token_data_start = pos_in_block * token_data_size |
| 1241 | fp8_bytes = cache_block[token_data_start : token_data_start + fp8_dim] |
| 1242 | fp8_vals = fp8_bytes.view(torch.float8_e4m3fn).float() |
| 1243 | |
| 1244 | scale_start = block_size * token_data_size + pos_in_block * scale_dim |
| 1245 | encoded_scales = cache_block[scale_start : scale_start + scale_dim] |
| 1246 | scales = torch.exp2(encoded_scales[:7].float() - 127.0) |
| 1247 | dequant = fp8_vals * scales.repeat_interleave(quant_block) |
| 1248 | |
| 1249 | bf16_start = token_data_start + fp8_dim |
| 1250 | bf16_bytes = cache_block[bf16_start : bf16_start + bf16_dim * 2] |
| 1251 | bf16_tail = bf16_bytes.view(torch.bfloat16) |
| 1252 | |
| 1253 | out[req_id, offset + i, :fp8_dim] = dequant |
| 1254 | out[req_id, offset + i, fp8_dim:] = bf16_tail |
| 1255 | |
| 1256 | |
| 1257 | @pytest.mark.parametrize( |
| 1258 | ("seq_lens_host", "gather_lens_host", "offset"), |
| 1259 | [ |
| 1260 | ([9, 23, 7], None, 0), |
| 1261 | ([19, 8, 257], [6, 8, 129], 5), |
| 1262 | ], |
| 1263 | ) |
| 1264 | def test_dequantize_and_gather_k_cache( |
| 1265 | seq_lens_host: list[int], |
| 1266 | gather_lens_host: list[int] | None, |
| 1267 | offset: int, |
| 1268 | ): |
| 1269 | block_size = 64 |
| 1270 | head_dim = 512 |
| 1271 | nope_dim = 448 |
| 1272 | scale_dim = 8 |
| 1273 | head_bytes = nope_dim + (head_dim - nope_dim) * 2 + scale_dim |
| 1274 | device = "cuda" |
| 1275 | num_reqs = len(seq_lens_host) |
| 1276 | num_tokens = sum(seq_lens_host) |
| 1277 | max_gather_len = max(gather_lens_host or seq_lens_host) |
| 1278 | max_blocks_per_seq = math.ceil(max(seq_lens_host) / block_size) |
| 1279 | num_blocks = sum(math.ceil(seq_len / block_size) for seq_len in seq_lens_host) |
| 1280 | |
| 1281 | compressed_kv = torch.randn( |
| 1282 | num_tokens, head_dim, dtype=torch.bfloat16, device=device |
| 1283 | ) |
| 1284 | |
| 1285 | # Randomize physical pages so the test covers block-table translation. |
| 1286 | # Keep padded block-table entries invalid to catch accidental reads. |
| 1287 | physical_blocks = torch.randperm(num_blocks, device=device) |
| 1288 | block_table = torch.full( |
| 1289 | (num_reqs, max_blocks_per_seq), int(-1e6), dtype=torch.int32, device=device |
| 1290 | ) |
| 1291 | start = 0 |
| 1292 | for req_id, seq_len in enumerate(seq_lens_host): |
| 1293 | num_req_blocks = math.ceil(seq_len / block_size) |
| 1294 | req_blocks = physical_blocks[start : start + num_req_blocks] |
| 1295 | block_table[req_id, :num_req_blocks] = req_blocks |
| 1296 | start += num_req_blocks |
| 1297 | |
| 1298 | # Build slot_mapping for quantize_and_insert_k_cache. |
| 1299 | slot_mapping = torch.empty(num_tokens, dtype=torch.int64, device=device) |
| 1300 | start = 0 |
| 1301 | for req_id, seq_len in enumerate(seq_lens_host): |
| 1302 | logical_pos = torch.arange(seq_len, dtype=torch.int64, device=device) |
| 1303 | block_idx = block_table[req_id, logical_pos // block_size].to(torch.int64) |
| 1304 | token_slots = block_idx * block_size + logical_pos % block_size |
| 1305 | slot_mapping[start : start + seq_len] = token_slots |
| 1306 | start += seq_len |
| 1307 | |
| 1308 | # Insert compressed K into the paged cache layout used by the gather op. |
| 1309 | k_cache = torch.empty( |
| 1310 | num_blocks, block_size, head_bytes, dtype=torch.uint8, device=device |
| 1311 | ) |
| 1312 | k_cache_2d = k_cache.view(num_blocks, -1) |
| 1313 | quantize_and_insert_k_cache(compressed_kv, k_cache_2d, slot_mapping, block_size) |
| 1314 | |
| 1315 | out_shape = (num_reqs, offset + max_gather_len + 3, head_dim) |
| 1316 | ref_out = torch.empty(out_shape, dtype=torch.bfloat16, device=device) |
| 1317 | actual_out = torch.empty_like(ref_out) |
| 1318 | seq_lens = torch.tensor(seq_lens_host, dtype=torch.int32, device=device) |
| 1319 | gather_lens = ( |
| 1320 | torch.tensor(gather_lens_host, dtype=torch.int32, device=device) |
| 1321 | if gather_lens_host is not None |
| 1322 | else None |
| 1323 | ) |
| 1324 | |
| 1325 | # Compare production gather against a PyTorch reference for valid output rows. |
| 1326 | _dequantize_and_gather_k_cache_reference( |
| 1327 | ref_out, k_cache, seq_lens, gather_lens, block_table, block_size, offset |
| 1328 | ) |
| 1329 | dequantize_and_gather_k_cache( |
| 1330 | actual_out, k_cache, seq_lens, gather_lens, block_table, block_size, offset |
| 1331 | ) |
| 1332 | torch.accelerator.synchronize() |
| 1333 | |
| 1334 | # only check non-padded content |
| 1335 | for req_id, seq_len in enumerate(seq_lens_host): |
| 1336 | gather_len = ( |
| 1337 | gather_lens_host[req_id] if gather_lens_host is not None else seq_len |
| 1338 | ) |
| 1339 | actual = actual_out[req_id, offset : offset + gather_len] |
| 1340 | expected = ref_out[req_id, offset : offset + gather_len] |
| 1341 | torch.testing.assert_close(actual, expected, rtol=0, atol=0) |
| 1342 | |
| 1343 | |
| 1344 | # ── Test C: Indexer path ──────────────────────────────────────────────────── |
| 1345 | |
| 1346 | |
| 1347 | @pytest.mark.parametrize("num_tokens", [1, 4, 8, 17]) |
| 1348 | @pytest.mark.parametrize("block_size", [16, 64]) |
| 1349 | def test_indexer_quant_cache_roundtrip(num_tokens: int, block_size: int): |
| 1350 | """K → indexer_k_quant_and_cache → cp_gather_indexer_k_quant_cache |
| 1351 | → manual dequant → compare against original.""" |
| 1352 | HEAD_DIM = 128 |
| 1353 | QUANT_BLOCK_SIZE = 128 |
| 1354 | # cache_stride = head_dim + (head_dim * 4 / quant_block_size) = 128 + 4 = 132 |
| 1355 | CACHE_STRIDE = HEAD_DIM + HEAD_DIM * 4 // QUANT_BLOCK_SIZE |
| 1356 | |
| 1357 | num_blocks = (num_tokens + block_size - 1) // block_size + 1 |
| 1358 | device = "cuda" |
| 1359 | |
| 1360 | # Random K (simulates compressor output for indexer) |
| 1361 | k = torch.randn(num_tokens, HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 1362 | |
| 1363 | # ── Quant + insert ────────────────────────────────────────────────── |
| 1364 | kv_cache = torch.zeros( |
| 1365 | num_blocks, block_size, CACHE_STRIDE, dtype=torch.uint8, device=device |
| 1366 | ) |
| 1367 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 1368 | |
| 1369 | ops.indexer_k_quant_and_cache(k, kv_cache, slot_mapping, QUANT_BLOCK_SIZE, "ue8m0") |
| 1370 | |
| 1371 | # ── Gather ────────────────────────────────────────────────────────── |
| 1372 | max_blocks_per_seq = num_blocks |
| 1373 | block_table = torch.arange( |
| 1374 | max_blocks_per_seq, dtype=torch.int32, device=device |
| 1375 | ).unsqueeze(0) |
| 1376 | cu_seq_lens = torch.tensor([0, num_tokens], dtype=torch.int32, device=device) |
| 1377 | |
| 1378 | # dst_k: [total_seq_len, head_dim] as uint8 (raw FP8 bytes) |
| 1379 | dst_k = torch.zeros(num_tokens, HEAD_DIM, dtype=torch.uint8, device=device) |
| 1380 | # dst_scale: [total_seq_len, head_dim/quant_block*4] as uint8 (raw float32 bytes) |
| 1381 | num_scale_bytes = HEAD_DIM * 4 // QUANT_BLOCK_SIZE # 4 |
| 1382 | dst_scale = torch.zeros( |
| 1383 | num_tokens, num_scale_bytes, dtype=torch.uint8, device=device |
| 1384 | ) |
| 1385 | |
| 1386 | ops.cp_gather_indexer_k_quant_cache( |
| 1387 | kv_cache, dst_k, dst_scale, block_table, cu_seq_lens |
| 1388 | ) |
| 1389 | |
| 1390 | # ── Manual dequant ────────────────────────────────────────────────── |
| 1391 | k_fp8 = dst_k.view(torch.float8_e4m3fn).float() # [num_tokens, 128] |
| 1392 | scale = dst_scale.view(torch.float32) # [num_tokens, 1] |
| 1393 | k_recovered = k_fp8 * scale # [num_tokens, 128] |
| 1394 | |
| 1395 | # ── Compare ───────────────────────────────────────────────────────── |
| 1396 | diff = (k_recovered - k.float()).abs() |
| 1397 | k_abs = k.float().abs() |
| 1398 | |
| 1399 | for t in range(num_tokens): |
| 1400 | amax = k_abs[t].max().clamp(min=1e-4).item() |
| 1401 | # UE8M0: scale = 2^ceil(log2(amax / 448)) |
| 1402 | exponent = math.ceil(math.log2(amax / 448.0)) |
| 1403 | ue8m0_scale = 2.0**exponent |
| 1404 | # FP8 e4m3 (3-bit mantissa): worst-case error = 16 * scale |
| 1405 | max_allowed = 16.0 * ue8m0_scale |
| 1406 | token_diff = diff[t].max().item() |
| 1407 | assert token_diff <= max_allowed, ( |
| 1408 | f"Token {t} diff {token_diff} exceeds max_allowed " |
| 1409 | f"{max_allowed} (scale={ue8m0_scale})" |
| 1410 | ) |
| 1411 | |
| 1412 | |
| 1413 | def test_indexer_gather_accepts_upper_bound_output(): |
| 1414 | """Gather only exact cu_seq_lens even when dst is over-allocated.""" |
| 1415 | head_dim = 128 |
| 1416 | quant_block_size = 128 |
| 1417 | cache_stride = head_dim + head_dim * 4 // quant_block_size |
| 1418 | valid_tokens = 9 |
| 1419 | upper_bound_tokens = 13 |
| 1420 | block_size = 16 |
| 1421 | num_seqs = 3 |
| 1422 | num_blocks = num_seqs |
| 1423 | sentinel = 123 |
| 1424 | device = "cuda" |
| 1425 | |
| 1426 | k = torch.randn(valid_tokens, head_dim, dtype=torch.bfloat16, device=device) |
| 1427 | kv_cache = torch.zeros( |
| 1428 | num_blocks, block_size, cache_stride, dtype=torch.uint8, device=device |
| 1429 | ) |
| 1430 | slot_mapping = torch.tensor( |
| 1431 | [0, 1, 2, 16, 17, 18, 32, 33, 34], dtype=torch.int64, device=device |
| 1432 | ) |
| 1433 | ops.indexer_k_quant_and_cache(k, kv_cache, slot_mapping, quant_block_size, "ue8m0") |
| 1434 | |
| 1435 | block_table = torch.arange(num_blocks, dtype=torch.int32, device=device).unsqueeze( |
| 1436 | 1 |
| 1437 | ) |
| 1438 | cu_seq_lens = torch.tensor([0, 3, 6, 9], dtype=torch.int32, device=device) |
| 1439 | dst_k = torch.full( |
| 1440 | (upper_bound_tokens, head_dim), sentinel, dtype=torch.uint8, device=device |
| 1441 | ) |
| 1442 | num_scale_bytes = head_dim * 4 // quant_block_size |
| 1443 | dst_scale = torch.full( |
| 1444 | (upper_bound_tokens, num_scale_bytes), |
| 1445 | sentinel, |
| 1446 | dtype=torch.uint8, |
| 1447 | device=device, |
| 1448 | ) |
| 1449 | |
| 1450 | ops.cp_gather_indexer_k_quant_cache( |
| 1451 | kv_cache, dst_k, dst_scale, block_table, cu_seq_lens |
| 1452 | ) |
| 1453 | |
| 1454 | if current_platform.is_rocm(): |
| 1455 | triton_kv_cache = torch.zeros_like(kv_cache) |
| 1456 | indexer_k_quant_and_cache_triton( |
| 1457 | k, |
| 1458 | triton_kv_cache, |
| 1459 | slot_mapping, |
| 1460 | quant_block_size, |
| 1461 | "ue8m0", |
| 1462 | ) |
| 1463 | triton_dst_k = torch.full_like(dst_k, sentinel) |
| 1464 | triton_dst_scale = torch.full_like(dst_scale, sentinel) |
| 1465 | token_to_seq = torch.cat( |
| 1466 | ( |
| 1467 | torch.repeat_interleave( |
| 1468 | torch.arange(num_seqs, dtype=torch.int32, device=device), 3 |
| 1469 | ), |
| 1470 | torch.full( |
| 1471 | (upper_bound_tokens - valid_tokens,), |
| 1472 | -1, |
| 1473 | dtype=torch.int32, |
| 1474 | device=device, |
| 1475 | ), |
| 1476 | ) |
| 1477 | ) |
| 1478 | cp_gather_indexer_k_quant_cache_triton( |
| 1479 | triton_kv_cache, |
| 1480 | triton_dst_k.view(current_platform.fp8_dtype()), |
| 1481 | triton_dst_scale, |
| 1482 | block_table, |
| 1483 | cu_seq_lens, |
| 1484 | token_to_seq, |
| 1485 | ) |
| 1486 | torch.accelerator.synchronize() |
| 1487 | |
| 1488 | if current_platform.is_rocm(): |
| 1489 | triton_recovered = triton_dst_k[:valid_tokens].view( |
| 1490 | current_platform.fp8_dtype() |
| 1491 | ).float() * triton_dst_scale[:valid_tokens].view(torch.float32) |
| 1492 | triton_error = (triton_recovered - k.float()).abs().amax(dim=1) |
| 1493 | max_triton_error = ( |
| 1494 | 16.0 * triton_dst_scale[:valid_tokens].view(torch.float32).flatten() |
| 1495 | ) |
| 1496 | assert torch.all(triton_error <= max_triton_error) |
| 1497 | assert torch.all(triton_dst_k[valid_tokens:] == sentinel) |
| 1498 | assert torch.all(triton_dst_scale[valid_tokens:] == sentinel) |
| 1499 | k_recovered = dst_k[:valid_tokens].view(torch.float8_e4m3fn).float() * dst_scale[ |
| 1500 | :valid_tokens |
| 1501 | ].view(torch.float32) |
| 1502 | diff = (k_recovered - k.float()).abs() |
| 1503 | max_allowed = (16.0 * dst_scale[:valid_tokens].view(torch.float32).max()).item() |
| 1504 | assert diff.max().item() <= max_allowed |
| 1505 | assert torch.all(dst_k[valid_tokens:] == sentinel) |
| 1506 | assert torch.all(dst_scale[valid_tokens:] == sentinel) |
| 1507 | |
| 1508 | |
| 1509 | # ── Test D: DeepseekV4 attention with values at different magnitudes ─────────── |
| 1510 | |
| 1511 | |
| 1512 | def test_deepseek_v4_quant_magnitude_range(): |
| 1513 | """Test that quantization handles a range of magnitudes correctly.""" |
| 1514 | HEAD_DIM = 512 |
| 1515 | NOPE_DIM = 448 |
| 1516 | HEAD_BYTES = 584 |
| 1517 | block_size = 16 |
| 1518 | num_tokens = 4 |
| 1519 | num_blocks = 2 |
| 1520 | device = "cuda" |
| 1521 | |
| 1522 | # Create inputs with varying magnitudes: small, medium, large |
| 1523 | compressed_kv = torch.zeros( |
| 1524 | num_tokens, HEAD_DIM, dtype=torch.bfloat16, device=device |
| 1525 | ) |
| 1526 | compressed_kv[0] = 0.001 # very small |
| 1527 | compressed_kv[1] = 1.0 # unit scale |
| 1528 | compressed_kv[2] = 100.0 # large |
| 1529 | compressed_kv[3] = torch.randn(HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 1530 | |
| 1531 | k_cache = torch.zeros( |
| 1532 | num_blocks, block_size, HEAD_BYTES, dtype=torch.uint8, device=device |
| 1533 | ) |
| 1534 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 1535 | |
| 1536 | quantize_and_insert_k_cache( |
| 1537 | compressed_kv, k_cache.view(num_blocks, -1), slot_mapping, block_size |
| 1538 | ) |
| 1539 | |
| 1540 | out = torch.zeros(1, num_tokens, HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 1541 | seq_lens = torch.tensor([num_tokens], dtype=torch.int32, device=device) |
| 1542 | block_table = torch.arange(num_blocks, dtype=torch.int32, device=device).unsqueeze( |
| 1543 | 0 |
| 1544 | ) |
| 1545 | |
| 1546 | dequantize_and_gather_k_cache( |
| 1547 | out, k_cache, seq_lens, None, block_table, block_size, offset=0 |
| 1548 | ) |
| 1549 | |
| 1550 | recovered = out[0, :num_tokens] |
| 1551 | |
| 1552 | # RoPE portion must be exact |
| 1553 | rope_diff = (recovered[:, NOPE_DIM:] - compressed_kv[:, NOPE_DIM:]).abs().max() |
| 1554 | assert rope_diff.item() == 0.0, f"RoPE diff {rope_diff.item()}" |
| 1555 | |
| 1556 | # NoPE: relative error should be reasonable |
| 1557 | for t in range(num_tokens): |
| 1558 | orig = compressed_kv[t, :NOPE_DIM].float() |
| 1559 | recv = recovered[t, :NOPE_DIM].float() |
| 1560 | abs_diff = (recv - orig).abs().max().item() |
| 1561 | magnitude = orig.abs().max().item() |
| 1562 | if magnitude > 0.01: |
| 1563 | rel_err = abs_diff / magnitude |
| 1564 | assert rel_err < 0.15, ( |
| 1565 | f"Token {t}: rel_err={rel_err:.4f}, abs_diff={abs_diff:.6f}, " |
| 1566 | f"magnitude={magnitude:.4f}" |
| 1567 | ) |
| 1568 | |
| 1569 | |
| 1570 | # ── Test E: Indexer fused K-cache insert (Triton kernels) ──────────────────── |
| 1571 | # |
| 1572 | # Both kernels share the same Triton signature; use_fp4 selects between them. |
| 1573 | # Full pipeline: state-cache gather → softmax-weighted compress → RMSNorm → |
| 1574 | # GPT-J RoPE → quant (MXFP4 or FP8) → paged cache insert. |
| 1575 | |
| 1576 | |
| 1577 | def _reference_kv_compress_norm_rope( |
| 1578 | state_cache: torch.Tensor, |
| 1579 | block_table: torch.Tensor, |
| 1580 | positions: torch.Tensor, |
| 1581 | rms_weight: torch.Tensor, |
| 1582 | cos_sin_cache: torch.Tensor, |
| 1583 | compress_ratio: int = 1, |
| 1584 | overlap: int = 0, |
| 1585 | use_fp4: bool = False, |
| 1586 | rms_eps: float = 1e-6, |
| 1587 | fp8_max: float = 448.0, |
| 1588 | return_full_cache: bool = False, |
| 1589 | ): |
| 1590 | """Compress → RMSNorm → GPT-J RoPE → quantize. |
| 1591 | |
| 1592 | Gathers (1+overlap)*compress_ratio state entries per output token, applies |
| 1593 | per-element softmax over the scores, and computes the weighted kv sum. |
| 1594 | Returns (quantized_values, scale) matching the kernel's output layout. |
| 1595 | """ |
| 1596 | device = state_cache.device |
| 1597 | head_dim = rms_weight.shape[0] |
| 1598 | rope_dim = cos_sin_cache.shape[-1] |
| 1599 | state_block_size = state_cache.shape[1] |
| 1600 | state_width = state_cache.shape[-1] // 2 |
| 1601 | nope_dim = head_dim - rope_dim |
| 1602 | total = (1 + overlap) * compress_ratio |
| 1603 | results = [] |
| 1604 | for pos in positions.tolist(): |
| 1605 | src = torch.arange(pos - total + 1, pos + 1, dtype=torch.int64, device=device) |
| 1606 | valid = src >= 0 |
| 1607 | idx = src.clamp(min=0) |
| 1608 | pages = block_table[0, idx // state_block_size] |
| 1609 | offsets = idx % state_block_size |
| 1610 | raw = state_cache[pages, offsets].float() # [total, state_dim] |
| 1611 | |
| 1612 | # Group 0 (tokens 0..cr-1): kv[:H], score[SW:SW+H] |
| 1613 | # Group 1 (tokens cr..2cr-1): kv[H:2H], score[SW+H:SW+2H] |
| 1614 | if overlap: |
| 1615 | sw = state_width |
| 1616 | g0_kv = raw[:compress_ratio, :head_dim] |
| 1617 | g1_kv = raw[compress_ratio:, head_dim : 2 * head_dim] |
| 1618 | g0_scores = raw[:compress_ratio, sw : sw + head_dim] |
| 1619 | g1_scores = raw[compress_ratio:, sw + head_dim : sw + 2 * head_dim] |
| 1620 | kv = torch.cat([g0_kv, g1_kv]) |
| 1621 | scores = torch.cat([g0_scores, g1_scores]) |
| 1622 | else: |
| 1623 | kv = raw[:, :head_dim] |
| 1624 | scores = raw[:, state_width : state_width + head_dim] |
| 1625 | |
| 1626 | scores[~valid] = float("-inf") |
| 1627 | kv[~valid] = 0.0 |
| 1628 | weights = torch.softmax(scores, dim=0) |
| 1629 | compressed = (kv * weights).sum(dim=0) # [H] |
| 1630 | var = (compressed * compressed).mean() |
| 1631 | normed = compressed * torch.rsqrt(var + rms_eps) * rms_weight.float() |
| 1632 | compressed_pos = (pos // compress_ratio) * compress_ratio |
| 1633 | cos, sin = cos_sin_cache[compressed_pos].float().chunk(2) |
| 1634 | nope, rope = normed.split([nope_dim, rope_dim]) |
| 1635 | rope = torch.stack( |
| 1636 | [rope[0::2] * cos - rope[1::2] * sin, rope[1::2] * cos + rope[0::2] * sin], |
| 1637 | dim=-1, |
| 1638 | ).reshape(rope_dim) |
| 1639 | results.append(torch.cat([nope, rope]).to(state_cache.dtype)) |
| 1640 | result = torch.stack(results) |
| 1641 | |
| 1642 | if return_full_cache: |
| 1643 | # Contiguous 512-wide bf16 row (nope unrotated + rope rotated), matching |
| 1644 | # the FlashInfer full-cache layout before any per-tensor fp8 quant. The |
| 1645 | # kernel rounds the fp32 result to bf16 once at the store. |
| 1646 | return result.to(torch.bfloat16) |
| 1647 | |
| 1648 | if use_fp4: |
| 1649 | return quantize_to_mxfp4(result) |
| 1650 | else: |
| 1651 | pairs = [ |
| 1652 | _ue8m0_reference(result[t], head_dim, fp8_max) for t in range(len(result)) |
| 1653 | ] |
| 1654 | quants, scales = zip(*pairs) |
| 1655 | return torch.stack(quants), torch.cat(scales) |
| 1656 | |
| 1657 | |
| 1658 | @pytest.mark.parametrize("num_tokens", [1, 7, 32]) |
| 1659 | @pytest.mark.parametrize("kv_block_size", [16, 32]) |
| 1660 | @pytest.mark.parametrize( |
| 1661 | "use_fp4", |
| 1662 | [ |
| 1663 | False, |
| 1664 | pytest.param( |
| 1665 | True, |
| 1666 | marks=pytest.mark.skipif( |
| 1667 | not ( |
| 1668 | current_platform.is_cuda() |
| 1669 | and current_platform.is_device_capability_family(100) |
| 1670 | ), |
| 1671 | reason="MXFP4 indexer cache requires an SM100-family GPU", |
| 1672 | ), |
| 1673 | ), |
| 1674 | ], |
| 1675 | ) |
| 1676 | def test_fused_kv_insert_indexer(num_tokens: int, kv_block_size: int, use_fp4: bool): |
| 1677 | """Fused K compress+norm+rope+quant+insert for the indexer KV cache.""" |
| 1678 | HEAD_DIM = 128 |
| 1679 | ROPE_DIM = 64 |
| 1680 | BLOCK_SIZE = 16 |
| 1681 | RMS_EPS = 1e-6 |
| 1682 | FP8_MAX = 448.0 |
| 1683 | |
| 1684 | device = "cuda" |
| 1685 | torch.manual_seed(42) |
| 1686 | compress_ratio = 4 |
| 1687 | |
| 1688 | if use_fp4: |
| 1689 | TOKEN_STRIDE = HEAD_DIM // 2 # packed nibbles: 64 bytes |
| 1690 | SCALE_DIM = HEAD_DIM // 32 # ue8m0 bytes: 4 |
| 1691 | QUANT_BLOCK = 32 |
| 1692 | kernel = _fused_kv_compress_norm_rope_insert_indexer_mxfp4_attn |
| 1693 | else: |
| 1694 | TOKEN_STRIDE = HEAD_DIM # FP8 bytes: 128 |
| 1695 | SCALE_DIM = 4 # 1 float32: 4 bytes |
| 1696 | QUANT_BLOCK = HEAD_DIM |
| 1697 | kernel = _fused_kv_compress_norm_rope_insert_indexer_attn |
| 1698 | |
| 1699 | # overlap=1 whenever compress_ratio==4, matching DeepseekCompressor logic. |
| 1700 | overlap = 1 if compress_ratio == 4 else 0 |
| 1701 | coff = 1 + overlap # multiplier for state_dim per entry |
| 1702 | |
| 1703 | num_pages = (compress_ratio * num_tokens - 1) // BLOCK_SIZE + 2 |
| 1704 | state_cache = torch.randn( |
| 1705 | num_pages, |
| 1706 | BLOCK_SIZE, |
| 1707 | 2 * coff * HEAD_DIM, # kv_state + score_state, each coff*HEAD_DIM wide |
| 1708 | dtype=torch.bfloat16, |
| 1709 | device=device, |
| 1710 | ) |
| 1711 | block_table = torch.arange(num_pages, dtype=torch.int32, device=device).unsqueeze(0) |
| 1712 | token_to_req = torch.zeros(num_tokens, dtype=torch.int32, device=device) |
| 1713 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 1714 | positions = torch.arange( |
| 1715 | compress_ratio - 1, |
| 1716 | compress_ratio * num_tokens, |
| 1717 | compress_ratio, |
| 1718 | dtype=torch.int64, |
| 1719 | device=device, |
| 1720 | ) |
| 1721 | rms_weight = torch.randn(HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 1722 | cos_sin_cache = torch.randn(compress_ratio * num_tokens, ROPE_DIM, device=device) |
| 1723 | |
| 1724 | kv_n_blocks = (num_tokens + kv_block_size - 1) // kv_block_size + 1 |
| 1725 | kv_cache = torch.zeros( |
| 1726 | kv_n_blocks, |
| 1727 | kv_block_size * (TOKEN_STRIDE + SCALE_DIM), |
| 1728 | dtype=torch.uint8, |
| 1729 | device=device, |
| 1730 | ) |
| 1731 | |
| 1732 | kernel[(num_tokens,)]( |
| 1733 | state_cache, |
| 1734 | state_cache.stride(0), |
| 1735 | state_cache.stride(1), |
| 1736 | token_to_req, |
| 1737 | positions, |
| 1738 | slot_mapping, |
| 1739 | block_table, |
| 1740 | block_table.stride(0), |
| 1741 | BLOCK_SIZE, |
| 1742 | rms_weight, |
| 1743 | RMS_EPS, |
| 1744 | cos_sin_cache, |
| 1745 | cos_sin_cache.stride(0), |
| 1746 | kv_cache, |
| 1747 | slot_mapping, |
| 1748 | kv_block_size, |
| 1749 | HEAD_SIZE=HEAD_DIM, |
| 1750 | TRITON_BLOCK_SIZE=HEAD_DIM, |
| 1751 | STATE_WIDTH=coff * HEAD_DIM, |
| 1752 | COMPRESS_RATIO=compress_ratio, |
| 1753 | OVERLAP=overlap, |
| 1754 | ROPE_HEAD_DIM=ROPE_DIM, |
| 1755 | FP8_MAX=FP8_MAX, |
| 1756 | QUANT_BLOCK=QUANT_BLOCK, |
| 1757 | TOKEN_STRIDE=TOKEN_STRIDE, |
| 1758 | SCALE_DIM=SCALE_DIM, |
| 1759 | KV_BLOCK_STRIDE=kv_cache.stride(0), |
| 1760 | num_warps=1, |
| 1761 | ) |
| 1762 | |
| 1763 | k_quant, scale = _reference_kv_compress_norm_rope( |
| 1764 | state_cache, |
| 1765 | block_table, |
| 1766 | positions, |
| 1767 | rms_weight, |
| 1768 | cos_sin_cache, |
| 1769 | compress_ratio, |
| 1770 | overlap, |
| 1771 | use_fp4, |
| 1772 | rms_eps=RMS_EPS, |
| 1773 | fp8_max=FP8_MAX, |
| 1774 | ) |
| 1775 | |
| 1776 | if use_fp4: |
| 1777 | for i in range(num_tokens): |
| 1778 | blk, pos = i // kv_block_size, i % kv_block_size |
| 1779 | val_off = pos * TOKEN_STRIDE |
| 1780 | fp4_actual = kv_cache[blk, val_off : val_off + TOKEN_STRIDE] |
| 1781 | assert torch.equal(k_quant[i], fp4_actual), ( |
| 1782 | f"token {i}: packed nibbles differ, " |
| 1783 | f"{(k_quant[i] != fp4_actual).sum()} " |
| 1784 | f"/ {TOKEN_STRIDE}" |
| 1785 | ) |
| 1786 | |
| 1787 | scale_off = kv_block_size * TOKEN_STRIDE + pos * SCALE_DIM |
| 1788 | scale_actual = kv_cache[blk, scale_off : scale_off + SCALE_DIM] |
| 1789 | assert torch.equal(scale_actual, scale[i]), ( |
| 1790 | f"token {i}: ue8m0 {scale_actual.tolist()} != {scale[i].tolist()}" |
| 1791 | ) |
| 1792 | |
| 1793 | else: |
| 1794 | k_quant = k_quant.view(torch.uint8) |
| 1795 | for i in range(num_tokens): |
| 1796 | blk, pos = i // kv_block_size, i % kv_block_size |
| 1797 | val_off = pos * TOKEN_STRIDE |
| 1798 | assert torch.equal( |
| 1799 | k_quant[i], kv_cache[blk, val_off : val_off + TOKEN_STRIDE] |
| 1800 | ), f"token {i}: FP8 bytes differ" |
| 1801 | |
| 1802 | scale_off = kv_block_size * TOKEN_STRIDE + pos * SCALE_DIM |
| 1803 | actual_scale = kv_cache[blk, scale_off : scale_off + SCALE_DIM].view( |
| 1804 | torch.float32 |
| 1805 | ) |
| 1806 | assert torch.equal(actual_scale, scale[i : i + 1]), ( |
| 1807 | f"token {i}: scale {actual_scale.item()} != {scale[i].item()}" |
| 1808 | ) |
| 1809 | |
| 1810 | |
| 1811 | @pytest.mark.parametrize("compress_ratio", [4, 128]) |
| 1812 | @pytest.mark.parametrize("store_fp8", [False, True]) |
| 1813 | def test_cutedsl_full_cache_store(compress_ratio: int, store_fp8: bool): |
| 1814 | """CuTeDSL compressor full-cache (FlashInfer) store parity for head=512. |
| 1815 | |
| 1816 | Exercises the contiguous bf16 / per-tensor fp8 store branch of both the C4 |
| 1817 | fused kernel and the C128 split kernel against the PyTorch reference. |
| 1818 | """ |
| 1819 | cutedsl = pytest.importorskip("cutlass") # noqa: F841 |
| 1820 | from vllm.models.deepseek_v4.nvidia.ops.sparse_attn_compress_cutedsl import ( |
| 1821 | fused_kv_compress_norm_rope_insert_sparse_attn_cutedsl, |
| 1822 | split_kv_compress_norm_rope_insert_sparse_attn_cutedsl, |
| 1823 | ) |
| 1824 | |
| 1825 | HEAD_DIM = 512 |
| 1826 | ROPE_DIM = 64 |
| 1827 | RMS_EPS = 1e-6 |
| 1828 | FP8_MAX = 448.0 |
| 1829 | # C128 compress (Block8 kernel) requires state-cache block_size=8; C4 uses 16. |
| 1830 | BLOCK_SIZE = 8 if compress_ratio == 128 else 16 |
| 1831 | KV_BLOCK_SIZE = 64 |
| 1832 | device = "cuda" |
| 1833 | torch.manual_seed(7) |
| 1834 | |
| 1835 | overlap = 1 if compress_ratio == 4 else 0 |
| 1836 | coff = 1 + overlap |
| 1837 | num_tokens = 8 |
| 1838 | |
| 1839 | num_pages = (compress_ratio * num_tokens - 1) // BLOCK_SIZE + 2 |
| 1840 | # The production CompressorStateCache is fp32. |
| 1841 | state_cache = torch.randn( |
| 1842 | num_pages, BLOCK_SIZE, 2 * coff * HEAD_DIM, dtype=torch.float32, device=device |
| 1843 | ) |
| 1844 | block_table = torch.arange(num_pages, dtype=torch.int32, device=device).unsqueeze(0) |
| 1845 | token_to_req = torch.zeros(num_tokens, dtype=torch.int32, device=device) |
| 1846 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 1847 | positions = torch.arange( |
| 1848 | compress_ratio - 1, |
| 1849 | compress_ratio * num_tokens, |
| 1850 | compress_ratio, |
| 1851 | dtype=torch.int64, |
| 1852 | device=device, |
| 1853 | ) |
| 1854 | rms_weight = torch.randn(HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 1855 | cos_sin_cache = torch.randn( |
| 1856 | compress_ratio * num_tokens, ROPE_DIM, dtype=torch.float32, device=device |
| 1857 | ) |
| 1858 | |
| 1859 | dtype = torch.float8_e4m3fn if store_fp8 else torch.bfloat16 |
| 1860 | kv_n_blocks = (num_tokens + KV_BLOCK_SIZE - 1) // KV_BLOCK_SIZE + 1 |
| 1861 | k_cache = torch.zeros( |
| 1862 | kv_n_blocks, KV_BLOCK_SIZE, HEAD_DIM, dtype=dtype, device=device |
| 1863 | ) |
| 1864 | fp8_scale = torch.tensor( |
| 1865 | [0.5 if store_fp8 else 1.0], dtype=torch.float32, device=device |
| 1866 | ) |
| 1867 | |
| 1868 | if compress_ratio == 4: |
| 1869 | fused_kv_compress_norm_rope_insert_sparse_attn_cutedsl( |
| 1870 | state_cache, |
| 1871 | token_to_req, |
| 1872 | positions, |
| 1873 | slot_mapping, |
| 1874 | block_table, |
| 1875 | BLOCK_SIZE, |
| 1876 | rms_weight, |
| 1877 | RMS_EPS, |
| 1878 | cos_sin_cache, |
| 1879 | k_cache, |
| 1880 | slot_mapping, |
| 1881 | KV_BLOCK_SIZE, |
| 1882 | k_cache.stride(0), |
| 1883 | head_size=HEAD_DIM, |
| 1884 | state_width=coff * HEAD_DIM, |
| 1885 | rope_head_dim=ROPE_DIM, |
| 1886 | fp8_max=FP8_MAX, |
| 1887 | quant_block=64, |
| 1888 | token_stride=576, |
| 1889 | scale_dim=8, |
| 1890 | compress_ratio=compress_ratio, |
| 1891 | overlap=True, |
| 1892 | store_full_kv=True, |
| 1893 | store_full_fp8=store_fp8, |
| 1894 | fp8_scale=fp8_scale, |
| 1895 | ) |
| 1896 | else: |
| 1897 | compressed_kv = torch.empty( |
| 1898 | (num_tokens, HEAD_DIM), dtype=torch.float32, device=device |
| 1899 | ) |
| 1900 | split_kv_compress_norm_rope_insert_sparse_attn_cutedsl( |
| 1901 | state_cache, |
| 1902 | token_to_req, |
| 1903 | positions, |
| 1904 | slot_mapping, |
| 1905 | block_table, |
| 1906 | BLOCK_SIZE, |
| 1907 | compressed_kv, |
| 1908 | rms_weight, |
| 1909 | RMS_EPS, |
| 1910 | cos_sin_cache, |
| 1911 | k_cache, |
| 1912 | slot_mapping, |
| 1913 | KV_BLOCK_SIZE, |
| 1914 | k_cache.stride(0), |
| 1915 | head_size=HEAD_DIM, |
| 1916 | state_width=coff * HEAD_DIM, |
| 1917 | rope_head_dim=ROPE_DIM, |
| 1918 | fp8_max=FP8_MAX, |
| 1919 | quant_block=64, |
| 1920 | token_stride=576, |
| 1921 | scale_dim=8, |
| 1922 | compress_ratio=compress_ratio, |
| 1923 | overlap=bool(overlap), |
| 1924 | store_full_kv=True, |
| 1925 | store_full_fp8=store_fp8, |
| 1926 | fp8_scale=fp8_scale, |
| 1927 | ) |
| 1928 | |
| 1929 | ref = _reference_kv_compress_norm_rope( |
| 1930 | state_cache, |
| 1931 | block_table, |
| 1932 | positions, |
| 1933 | rms_weight, |
| 1934 | cos_sin_cache, |
| 1935 | compress_ratio, |
| 1936 | overlap, |
| 1937 | rms_eps=RMS_EPS, |
| 1938 | return_full_cache=True, |
| 1939 | ) # [num_tokens, HEAD_DIM] bf16 |
| 1940 | |
| 1941 | actual = torch.stack( |
| 1942 | [k_cache[i // KV_BLOCK_SIZE, i % KV_BLOCK_SIZE] for i in range(num_tokens)] |
| 1943 | ) |
| 1944 | if store_fp8: |
| 1945 | ref_fp8 = torch.clamp(ref.float() / fp8_scale, -FP8_MAX, FP8_MAX).to( |
| 1946 | torch.float8_e4m3fn |
| 1947 | ) |
| 1948 | torch.testing.assert_close(actual.float(), ref_fp8.float(), rtol=0.0, atol=0.3) |
| 1949 | else: |
| 1950 | torch.testing.assert_close(actual.float(), ref.float(), rtol=3e-2, atol=3e-2) |
| 1951 | |
| 1952 | |
| 1953 | # ── Test F: DeepseekV4 Attention two-stage split compressor (Triton) ───────── |
| 1954 | # |
| 1955 | # Same full pipeline as Test E (state-cache gather -> softmax-weighted compress |
| 1956 | # -> RMSNorm -> GPT-J RoPE -> quant -> paged insert), but for the head=512 |
| 1957 | # fp8_ds_mla layout via the two-stage split |
| 1958 | |
| 1959 | |
| 1960 | @pytest.mark.skipif( |
| 1961 | not current_platform.is_rocm(), |
| 1962 | reason="two-stage split compressor is only enabled for ROCm at the moment", |
| 1963 | ) |
| 1964 | @pytest.mark.parametrize("num_tokens", [1, 4, 8, 17]) |
| 1965 | @pytest.mark.parametrize("kv_block_size", [16, 64]) |
| 1966 | def test_fused_kv_insert_split(num_tokens: int, kv_block_size: int): |
| 1967 | """Two-stage split compress+norm+rope+quant+insert for the head=512 KV cache.""" |
| 1968 | HEAD_DIM = 512 |
| 1969 | NOPE_DIM = 448 |
| 1970 | ROPE_DIM = 64 |
| 1971 | HEAD_BYTES = 584 # 448 fp8 + 128 bf16 + 8 uint8 scale |
| 1972 | RMS_EPS = 1e-6 |
| 1973 | FP8_MAX = 448.0 |
| 1974 | QUANT_BLOCK = 64 |
| 1975 | TOKEN_STRIDE = 576 |
| 1976 | SCALE_DIM = 8 |
| 1977 | STATE_BLOCK_SIZE = 8 # CompressorStateCache block_size for cr=128 |
| 1978 | |
| 1979 | device = "cuda" |
| 1980 | torch.manual_seed(42) |
| 1981 | compress_ratio = 128 |
| 1982 | overlap = 0 # no overlap for cr=128 |
| 1983 | coff = 1 + overlap |
| 1984 | |
| 1985 | num_pages = (compress_ratio * num_tokens - 1) // STATE_BLOCK_SIZE + 2 |
| 1986 | state_cache = torch.randn( |
| 1987 | num_pages, |
| 1988 | STATE_BLOCK_SIZE, |
| 1989 | 2 * coff * HEAD_DIM, # kv_state + score_state |
| 1990 | dtype=torch.float32, |
| 1991 | device=device, |
| 1992 | ) |
| 1993 | block_table = torch.arange(num_pages, dtype=torch.int32, device=device).unsqueeze(0) |
| 1994 | token_to_req = torch.zeros(num_tokens, dtype=torch.int32, device=device) |
| 1995 | slot_mapping = torch.arange(num_tokens, dtype=torch.int64, device=device) |
| 1996 | positions = torch.arange( |
| 1997 | compress_ratio - 1, |
| 1998 | compress_ratio * num_tokens, |
| 1999 | compress_ratio, |
| 2000 | dtype=torch.int64, |
| 2001 | device=device, |
| 2002 | ) |
| 2003 | rms_weight = torch.randn(HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 2004 | cos_sin_cache = torch.randn( |
| 2005 | compress_ratio * num_tokens, ROPE_DIM, dtype=torch.float32, device=device |
| 2006 | ) |
| 2007 | |
| 2008 | kv_n_blocks = (num_tokens + kv_block_size - 1) // kv_block_size + 1 |
| 2009 | kv_cache = torch.zeros( |
| 2010 | kv_n_blocks, kv_block_size, HEAD_BYTES, dtype=torch.uint8, device=device |
| 2011 | ) |
| 2012 | compress_scratch = torch.empty( |
| 2013 | num_tokens, HEAD_DIM, dtype=torch.float32, device=device |
| 2014 | ) |
| 2015 | |
| 2016 | _launch_two_stage_sparse_attn_compressor( |
| 2017 | state_cache, |
| 2018 | token_to_req, |
| 2019 | positions, |
| 2020 | slot_mapping, |
| 2021 | block_table, |
| 2022 | STATE_BLOCK_SIZE, |
| 2023 | coff * HEAD_DIM, |
| 2024 | compress_ratio, |
| 2025 | cos_sin_cache, |
| 2026 | kv_cache, |
| 2027 | slot_mapping, |
| 2028 | rms_weight, |
| 2029 | RMS_EPS, |
| 2030 | QUANT_BLOCK, |
| 2031 | TOKEN_STRIDE, |
| 2032 | SCALE_DIM, |
| 2033 | HEAD_DIM, |
| 2034 | ROPE_DIM, |
| 2035 | num_tokens, |
| 2036 | compress_scratch, |
| 2037 | ) |
| 2038 | |
| 2039 | # PyTorch reference: compress -> RMSNorm -> GPT-J RoPE (pre-quant bf16 row). |
| 2040 | ref = _reference_kv_compress_norm_rope( |
| 2041 | state_cache, |
| 2042 | block_table, |
| 2043 | positions, |
| 2044 | rms_weight, |
| 2045 | cos_sin_cache, |
| 2046 | compress_ratio, |
| 2047 | overlap, |
| 2048 | rms_eps=RMS_EPS, |
| 2049 | fp8_max=FP8_MAX, |
| 2050 | return_full_cache=True, |
| 2051 | ) # [num_tokens, HEAD_DIM] bf16 |
| 2052 | |
| 2053 | # Dequant + gather the fp8_ds_mla cache back to bf16 (Test B op). |
| 2054 | out = torch.zeros(1, num_tokens, HEAD_DIM, dtype=torch.bfloat16, device=device) |
| 2055 | seq_lens = torch.tensor([num_tokens], dtype=torch.int32, device=device) |
| 2056 | gather_block_table = torch.arange( |
| 2057 | kv_n_blocks, dtype=torch.int32, device=device |
| 2058 | ).unsqueeze(0) |
| 2059 | dequantize_and_gather_k_cache( |
| 2060 | out, kv_cache, seq_lens, None, gather_block_table, kv_block_size, offset=0 |
| 2061 | ) |
| 2062 | recovered = out[0, :num_tokens] |
| 2063 | |
| 2064 | # NoPE (first 448): FP8 quantized, expect UE8M0 error (same bound as Test A). |
| 2065 | nope_diff = (recovered[:, :NOPE_DIM].float() - ref[:, :NOPE_DIM].float()).abs() |
| 2066 | for t in range(num_tokens): |
| 2067 | _, scales = _ue8m0_reference(ref[t, :NOPE_DIM].float(), QUANT_BLOCK, FP8_MAX) |
| 2068 | max_allowed = 16.0 * scales.max().item() |
| 2069 | token_diff = nope_diff[t].max().item() |
| 2070 | assert token_diff <= max_allowed, ( |
| 2071 | f"Token {t} nope diff {token_diff} exceeds max_allowed " |
| 2072 | f"{max_allowed} (scale={scales.max().item()})" |
| 2073 | ) |
| 2074 | |
| 2075 | # RoPE (last 64): stored as bf16. The kernel recomputes the rotation, so it |
| 2076 | # is bf16-close to the reference rather than bit-exact (cf. test_cutedsl). |
| 2077 | torch.testing.assert_close(recovered[:, NOPE_DIM:], ref[:, NOPE_DIM:]) |
| 2078 | |
| 2079 | |
| 2080 | @pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm stream coverage") |
| 2081 | def test_v41_rocm_csa2_full_pipeline(): |
| 2082 | """_forward_csa2_full: one fork, then K write and q-side after the join. |
| 2083 | |
| 2084 | The default chain must run first on the current stream, the compressor |
| 2085 | chain on aux 0 and the indexer weights projection on aux 1; the join must |
| 2086 | publish both cache writes before _produce_k / forward_q run. |
| 2087 | """ |
| 2088 | from vllm.forward_context import ForwardContext, override_forward_context |
| 2089 | from vllm.models.deepseek_v41.amd.rocm import DeepseekV41ROCMAiterMLAAttention |
| 2090 | from vllm.models.deepseek_v41.compressor import DeepseekCompressor |
| 2091 | |
| 2092 | torch.manual_seed(43) |
| 2093 | num_tokens = 19 |
| 2094 | hidden_states = torch.randn(num_tokens, 1024, device="cuda") |
| 2095 | positions = torch.arange(7, 26, device="cuda") |
| 2096 | qr_kv = torch.zeros(num_tokens, 1536, device="cuda") |
| 2097 | qr = torch.zeros(num_tokens, 4, device="cuda") |
| 2098 | kv = torch.zeros(num_tokens, 512, device="cuda") |
| 2099 | q = torch.zeros(num_tokens, 512, device="cuda") |
| 2100 | latent = torch.ones(num_tokens, 512, dtype=torch.bfloat16, device="cuda") |
| 2101 | weights = torch.randn(num_tokens, 32, device="cuda") |
| 2102 | index_q = torch.full((num_tokens, 32, 128), 2.0, device="cuda") |
| 2103 | index_q_scale = torch.full((num_tokens, 32), 3.0, device="cuda") |
| 2104 | index_weights_out = torch.full((num_tokens, 32), 4.0, device="cuda") |
| 2105 | main_cache = torch.full((1, 128, 584), 165, dtype=torch.uint8, device="cuda") |
| 2106 | index_cache = torch.full((1, 128, 132), 165, dtype=torch.uint8, device="cuda") |
| 2107 | |
| 2108 | order: list[str] = [] |
| 2109 | streams: dict[str, torch.cuda.Stream] = {} |
| 2110 | |
| 2111 | compressor = DeepseekCompressor.__new__(DeepseekCompressor) |
| 2112 | torch.nn.Module.__init__(compressor) |
| 2113 | compressor.fused_wkv_wgate = SimpleNamespace( |
| 2114 | weight=torch.randn(1024, 1024, device="cuda") |
| 2115 | ) |
| 2116 | |
| 2117 | def compressor_forward(kv_score, positions_): |
| 2118 | order.append("compressor") |
| 2119 | streams["compressor"] = torch.cuda.current_stream() |
| 2120 | return latent |
| 2121 | |
| 2122 | def compressor_insert(latent_, positions_, rotary_emb): |
| 2123 | order.append("insert_cache") |
| 2124 | streams["insert_cache"] = torch.cuda.current_stream() |
| 2125 | main_cache.fill_(5) |
| 2126 | |
| 2127 | compressor.forward = compressor_forward |
| 2128 | compressor.insert_cache = compressor_insert |
| 2129 | |
| 2130 | def weights_proj(hidden_states_): |
| 2131 | order.append("weights_proj") |
| 2132 | streams["weights_proj"] = torch.cuda.current_stream() |
| 2133 | return weights, None |
| 2134 | |
| 2135 | def produce_k(latent_, positions_, rotary_emb): |
| 2136 | order.append("produce_k") |
| 2137 | streams["produce_k"] = torch.cuda.current_stream() |
| 2138 | index_cache.fill_(7) |
| 2139 | |
| 2140 | def forward_q(qr_, qr_scale, indexer_weights, positions_, rotary_emb): |
| 2141 | order.append("forward_q") |
| 2142 | streams["forward_q"] = torch.cuda.current_stream() |
| 2143 | return index_q, index_q_scale, index_weights_out |
| 2144 | |
| 2145 | indexer = SimpleNamespace( |
| 2146 | weights_proj=weights_proj, |
| 2147 | _produce_k=produce_k, |
| 2148 | forward_q=forward_q, |
| 2149 | ) |
| 2150 | rotary = SimpleNamespace(cos_sin_cache=torch.randn(32, 64, device="cuda")) |
| 2151 | aux_streams = [torch.cuda.Stream(), torch.cuda.Stream()] |
| 2152 | |
| 2153 | def wq_b(qr_, qr_scale): |
| 2154 | order.append("wq_b") |
| 2155 | streams["wq_b"] = torch.cuda.current_stream() |
| 2156 | return q |
| 2157 | |
| 2158 | def qnorm_insert(q_, kv_, positions_, attn_metadata): |
| 2159 | order.append("qnorm_insert") |
| 2160 | streams["qnorm_insert"] = torch.cuda.current_stream() |
| 2161 | return q_ |
| 2162 | |
| 2163 | attention = SimpleNamespace( |
| 2164 | compressor=compressor, |
| 2165 | indexer=indexer, |
| 2166 | aux_stream_list=aux_streams, |
| 2167 | ln_events=[torch.cuda.Event() for _ in range(4)], |
| 2168 | rotary_emb=rotary, |
| 2169 | indexer_rotary_emb=rotary, |
| 2170 | n_local_heads=1, |
| 2171 | head_dim=512, |
| 2172 | _fused_wqa_wkv_gemm=lambda hidden_states_: qr_kv, |
| 2173 | _split_qkv_and_norm=lambda qr_kv_: (qr, None, kv), |
| 2174 | _wq_b_proj=wq_b, |
| 2175 | _fused_qnorm_rope_kv_insert=qnorm_insert, |
| 2176 | ) |
| 2177 | |
| 2178 | context = ForwardContext({}, {}, {}) |
| 2179 | with override_forward_context(context): |
| 2180 | result = DeepseekV41ROCMAiterMLAAttention._forward_csa2_full( |
| 2181 | attention, hidden_states, positions |
| 2182 | ) |
| 2183 | |
| 2184 | q_out, kv_out, index_q_out, index_q_scale_out, index_weights_out_out = result |
| 2185 | assert torch.equal( |
| 2186 | q_out, q.view(num_tokens, attention.n_local_heads, attention.head_dim) |
| 2187 | ) |
| 2188 | assert torch.equal(kv_out, kv) |
| 2189 | assert torch.equal(index_q_out, index_q) |
| 2190 | assert torch.equal(index_q_scale_out, index_q_scale) |
| 2191 | assert torch.equal(index_weights_out_out, index_weights_out) |
| 2192 | # Join publishes the aux-0 cache insert and orders the post-join K write. |
| 2193 | assert torch.equal(main_cache, torch.full_like(main_cache, 5)) |
| 2194 | assert torch.equal(index_cache, torch.full_like(index_cache, 7)) |
| 2195 | |
| 2196 | assert order == [ |
| 2197 | "wq_b", |
| 2198 | "qnorm_insert", |
| 2199 | "compressor", |
| 2200 | "insert_cache", |
| 2201 | "weights_proj", |
| 2202 | "produce_k", |
| 2203 | "forward_q", |
| 2204 | ] |
| 2205 | current = torch.cuda.current_stream() |
| 2206 | for name in ("wq_b", "qnorm_insert", "produce_k", "forward_q"): |
| 2207 | assert streams[name] == current, name |
| 2208 | for name in ("compressor", "insert_cache"): |
| 2209 | assert streams[name] == aux_streams[0], name |
| 2210 | assert streams["weights_proj"] == aux_streams[1] |
| 2211 | |
| 2212 | |
| 2213 | @pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm stream coverage") |
| 2214 | def test_v41_rocm_csa2_reindex_pipeline(): |
| 2215 | """_forward_csa2_reindex: serial projections, then SWA q path vs q-side.""" |
| 2216 | from vllm.forward_context import ForwardContext, override_forward_context |
| 2217 | from vllm.models.deepseek_v41.amd.rocm import DeepseekV41ROCMAiterMLAAttention |
| 2218 | |
| 2219 | torch.manual_seed(45) |
| 2220 | num_tokens = 19 |
| 2221 | hidden_states = torch.randn(num_tokens, 1024, device="cuda") |
| 2222 | positions = torch.arange(7, 26, device="cuda") |
| 2223 | qr_kv = torch.zeros(num_tokens, 1536, device="cuda") |
| 2224 | qr = torch.zeros(num_tokens, 4, device="cuda") |
| 2225 | kv = torch.zeros(num_tokens, 512, device="cuda") |
| 2226 | q = torch.zeros(num_tokens, 512, device="cuda") |
| 2227 | indexer_weights = torch.randn(num_tokens, 32, device="cuda") |
| 2228 | index_q = torch.full((num_tokens, 32, 128), 2.0, device="cuda") |
| 2229 | index_q_scale = torch.full((num_tokens, 32), 3.0, device="cuda") |
| 2230 | index_weights_out = torch.full((num_tokens, 32), 4.0, device="cuda") |
| 2231 | |
| 2232 | order: list[str] = [] |
| 2233 | streams: dict[str, torch.cuda.Stream] = {} |
| 2234 | |
| 2235 | def forward_q(qr_, qr_scale, indexer_weights_, positions_, rotary_emb): |
| 2236 | order.append("forward_q") |
| 2237 | streams["forward_q"] = torch.cuda.current_stream() |
| 2238 | return index_q, index_q_scale, index_weights_out |
| 2239 | |
| 2240 | indexer = SimpleNamespace(forward_q=forward_q) |
| 2241 | rotary = SimpleNamespace(cos_sin_cache=torch.randn(32, 64, device="cuda")) |
| 2242 | aux0 = torch.cuda.Stream() |
| 2243 | |
| 2244 | def wq_b(qr_, qr_scale): |
| 2245 | order.append("wq_b") |
| 2246 | streams["wq_b"] = torch.cuda.current_stream() |
| 2247 | return q |
| 2248 | |
| 2249 | def qnorm_insert(q_, kv_, positions_, attn_metadata): |
| 2250 | order.append("qnorm_insert") |
| 2251 | streams["qnorm_insert"] = torch.cuda.current_stream() |
| 2252 | return q_ |
| 2253 | |
| 2254 | attention = SimpleNamespace( |
| 2255 | indexer=indexer, |
| 2256 | aux_stream_list=[aux0], |
| 2257 | ln_events=[torch.cuda.Event() for _ in range(2)], |
| 2258 | indexer_rotary_emb=rotary, |
| 2259 | n_local_heads=1, |
| 2260 | head_dim=512, |
| 2261 | _run_parallel_input_projections=lambda hidden_states_: ( |
| 2262 | qr_kv, |
| 2263 | None, |
| 2264 | indexer_weights, |
| 2265 | ), |
| 2266 | _split_qkv_and_norm=lambda qr_kv_: (qr, None, kv), |
| 2267 | _wq_b_proj=wq_b, |
| 2268 | _fused_qnorm_rope_kv_insert=qnorm_insert, |
| 2269 | ) |
| 2270 | |
| 2271 | context = ForwardContext({}, {}, {}) |
| 2272 | with override_forward_context(context): |
| 2273 | result = DeepseekV41ROCMAiterMLAAttention._forward_csa2_reindex( |
| 2274 | attention, hidden_states, positions |
| 2275 | ) |
| 2276 | |
| 2277 | q_out, kv_out, index_q_out, index_q_scale_out, index_weights_out_out = result |
| 2278 | assert torch.equal( |
| 2279 | q_out, q.view(num_tokens, attention.n_local_heads, attention.head_dim) |
| 2280 | ) |
| 2281 | assert torch.equal(kv_out, kv) |
| 2282 | assert torch.equal(index_q_out, index_q) |
| 2283 | assert torch.equal(index_q_scale_out, index_q_scale) |
| 2284 | assert torch.equal(index_weights_out_out, index_weights_out) |
| 2285 | |
| 2286 | assert order == ["wq_b", "qnorm_insert", "forward_q"] |
| 2287 | current = torch.cuda.current_stream() |
| 2288 | for name in ("wq_b", "qnorm_insert"): |
| 2289 | assert streams[name] == current, name |
| 2290 | assert streams["forward_q"] == aux0 |
| 2291 | |
| 2292 | |
| 2293 | @pytest.mark.parametrize("quant_tuple", [False, True]) |
| 2294 | def test_v41_indexer_forward_q(quant_tuple): |
| 2295 | """forward_q runs wq_b then the fused RoPE/quant and unpacks both shapes.""" |
| 2296 | from unittest.mock import patch |
| 2297 | |
| 2298 | from vllm.models.deepseek_v41.attention import DeepseekV4Indexer |
| 2299 | |
| 2300 | num_tokens, n_head, head_dim = 5, 32, 128 |
| 2301 | qr = torch.randn(num_tokens, 1280) |
| 2302 | q = torch.randn(num_tokens, n_head * head_dim) |
| 2303 | positions = torch.arange(num_tokens) |
| 2304 | indexer_weights = torch.randn(num_tokens, n_head) |
| 2305 | rotary = SimpleNamespace(cos_sin_cache=torch.randn(32, 64)) |
| 2306 | fake = SimpleNamespace( |
| 2307 | n_head=n_head, |
| 2308 | head_dim=head_dim, |
| 2309 | softmax_scale=head_dim**-0.5, |
| 2310 | use_fp4_kv=False, |
| 2311 | indexer_weights_dtype=torch.float32, |
| 2312 | _wq_b_proj=lambda qr_, qr_scale: q, |
| 2313 | ) |
| 2314 | q_quant = torch.randn(num_tokens, n_head, head_dim) |
| 2315 | q_scale = torch.randn(num_tokens, n_head) |
| 2316 | weights_out = torch.randn(num_tokens, n_head) |
| 2317 | if quant_tuple: |
| 2318 | fused_return = ((q_quant, q_scale), weights_out) |
| 2319 | expected = (q_quant, q_scale, weights_out) |
| 2320 | else: |
| 2321 | fused_return = (q_quant, weights_out) |
| 2322 | expected = (q_quant, None, weights_out) |
| 2323 | |
| 2324 | with patch( |
| 2325 | "vllm.models.deepseek_v41.attention.fused_indexer_q_rope_quant", |
| 2326 | return_value=fused_return, |
| 2327 | ) as mock_fused: |
| 2328 | result = DeepseekV4Indexer.forward_q( |
| 2329 | fake, qr, None, indexer_weights, positions, rotary |
| 2330 | ) |
| 2331 | |
| 2332 | assert torch.equal(result[0], expected[0]) |
| 2333 | assert (result[1] is None) == (expected[1] is None) |
| 2334 | if expected[1] is not None: |
| 2335 | assert torch.equal(result[1], expected[1]) |
| 2336 | assert torch.equal(result[2], expected[2]) |
| 2337 | args = mock_fused.call_args.args |
| 2338 | assert args[0] is positions |
| 2339 | torch.testing.assert_close(args[1], q.view(-1, n_head, head_dim)) |
| 2340 | assert args[2] is rotary.cos_sin_cache |
| 2341 | assert args[3] is indexer_weights |
| 2342 | assert args[4] == head_dim**-0.5 |
| 2343 | assert args[5] == n_head**-0.5 |
| 2344 | assert mock_fused.call_args.kwargs == { |
| 2345 | "use_fp4": False, |
| 2346 | "weights_out_dtype": torch.float32, |
| 2347 | } |