| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | import torch.nn.functional as F |
| 7 | from einops import rearrange, repeat |
| 8 | |
| 9 | from tests.kernels.mamba.utils import selective_state_update_ref |
| 10 | from tests.kernels.utils import opcheck |
| 11 | from vllm import _custom_ops as ops # noqa: F401 |
| 12 | from vllm.model_executor.layers.mamba.ops.mamba_ssm import ( |
| 13 | selective_scan_fn, |
| 14 | selective_state_update, |
| 15 | ) |
| 16 | from vllm.platforms import current_platform |
| 17 | from vllm.utils.torch_utils import set_random_seed |
| 18 | from vllm.v1.attention.backends.utils import NULL_BLOCK_ID |
| 19 | |
| 20 | DEVICE = current_platform.device_type |
| 21 | |
| 22 | pytestmark = pytest.mark.skipif( |
| 23 | not ( |
| 24 | current_platform.is_cuda_alike() |
| 25 | or current_platform.is_xpu() |
| 26 | or current_platform.is_cpu() |
| 27 | ), |
| 28 | reason="mamba_ssm kernels require CUDA-alike, XPU, or CPU", |
| 29 | ) |
| 30 | |
| 31 | # selective_scan_fn is backed by the CUDA-only `ops.selective_scan_fwd` C++ op, |
| 32 | # so tests exercising it must be skipped on XPU. selective_state_update is |
| 33 | # pure Triton and runs on both CUDA-alike and XPU. |
| 34 | skip_unless_cuda_alike = pytest.mark.skipif( |
| 35 | not current_platform.is_cuda_alike(), |
| 36 | reason="selective_scan_fn uses CUDA-only custom op", |
| 37 | ) |
| 38 | |
| 39 | |
| 40 | def selective_scan_ref( |
| 41 | u, |
| 42 | delta, |
| 43 | A, |
| 44 | B, |
| 45 | C, |
| 46 | D=None, |
| 47 | z=None, |
| 48 | delta_bias=None, |
| 49 | delta_softplus=False, |
| 50 | return_last_state=False, |
| 51 | prev_state=None, |
| 52 | final_state_out=None, |
| 53 | ): |
| 54 | """u: r(B D L) |
| 55 | delta: r(B D L) |
| 56 | A: c(D N) or r(D N) |
| 57 | B: c(D N) or r(B N L) or r(B N 2L) or r(B G N L) or (B G N L) |
| 58 | C: c(D N) or r(B N L) or r(B N 2L) or r(B G N L) or (B G N L) |
| 59 | D: r(D) |
| 60 | z: r(B D L) |
| 61 | delta_bias: r(D), fp32 |
| 62 | prev_state: r(B D N), fp32 |
| 63 | |
| 64 | out: r(B D L) |
| 65 | last_state (optional): r(B D dstate) or c(B D dstate) |
| 66 | """ |
| 67 | dtype_in = u.dtype |
| 68 | u = u.float() |
| 69 | delta = delta.float() |
| 70 | if delta_bias is not None: |
| 71 | delta = delta + delta_bias[..., None].float() |
| 72 | if delta_softplus: |
| 73 | delta = F.softplus(delta) |
| 74 | batch, dim, dstate = u.shape[0], A.shape[0], A.shape[1] |
| 75 | is_variable_B = B.dim() >= 3 |
| 76 | is_variable_C = C.dim() >= 3 |
| 77 | B = B.float() |
| 78 | C = C.float() |
| 79 | x = A.new_zeros((batch, dim, dstate)) if prev_state is None else prev_state |
| 80 | ys = [] |
| 81 | deltaA = torch.exp(torch.einsum("bdl,dn->bdln", delta, A)) |
| 82 | if not is_variable_B: |
| 83 | deltaB_u = torch.einsum("bdl,dn,bdl->bdln", delta, B, u) |
| 84 | else: |
| 85 | if B.dim() == 3: |
| 86 | deltaB_u = torch.einsum("bdl,bnl,bdl->bdln", delta, B, u) |
| 87 | else: |
| 88 | B = repeat(B, "B G N L -> B (G H) N L", H=dim // B.shape[1]) |
| 89 | deltaB_u = torch.einsum("bdl,bdnl,bdl->bdln", delta, B, u) |
| 90 | if is_variable_C and C.dim() == 4: |
| 91 | C = repeat(C, "B G N L -> B (G H) N L", H=dim // C.shape[1]) |
| 92 | for i in range(u.shape[2]): |
| 93 | x = deltaA[:, :, i] * x + deltaB_u[:, :, i] |
| 94 | if not is_variable_C: |
| 95 | y = torch.einsum("bdn,dn->bd", x, C) |
| 96 | else: |
| 97 | if C.dim() == 3: |
| 98 | y = torch.einsum("bdn,bn->bd", x, C[:, :, i]) |
| 99 | else: |
| 100 | y = torch.einsum("bdn,bdn->bd", x, C[:, :, :, i]) |
| 101 | if i == u.shape[2] - 1: |
| 102 | if final_state_out is None: |
| 103 | final_state_out = x |
| 104 | else: |
| 105 | final_state_out.copy_(x) |
| 106 | ys.append(y) |
| 107 | y = torch.stack(ys, dim=2) # (batch dim L) |
| 108 | out = y if D is None else y + u * rearrange(D, "d -> d 1") |
| 109 | if z is not None: |
| 110 | out = out * F.silu(z) |
| 111 | out = out.to(dtype=dtype_in) |
| 112 | return out if not return_last_state else (out, final_state_out) |
| 113 | |
| 114 | |
| 115 | def selective_scan_opcheck_fn( |
| 116 | u, |
| 117 | delta, |
| 118 | A, |
| 119 | B, |
| 120 | C, |
| 121 | D=None, |
| 122 | z=None, |
| 123 | delta_bias=None, |
| 124 | delta_softplus=False, |
| 125 | cu_seq_len=None, |
| 126 | cache_indices=None, |
| 127 | has_initial_state=None, |
| 128 | ssm_states=None, |
| 129 | null_block_id=NULL_BLOCK_ID, |
| 130 | block_size=2048, |
| 131 | block_idx_first_scheduled_token=None, |
| 132 | block_idx_last_scheduled_token=None, |
| 133 | initial_state_idx=None, |
| 134 | cu_chunk_seqlen=None, |
| 135 | last_chunk_indices=None, |
| 136 | ): |
| 137 | """If return_last_state is True, returns (out, last_state) |
| 138 | last_state has shape (batch, dim, dstate). |
| 139 | """ |
| 140 | if u.stride(-1) != 1: |
| 141 | u = u.contiguous() |
| 142 | if delta.stride(-1) != 1: |
| 143 | delta = delta.contiguous() |
| 144 | if D is not None: |
| 145 | D = D.contiguous() |
| 146 | if B.stride(-1) != 1: |
| 147 | B = B.contiguous() |
| 148 | if C.stride(-1) != 1: |
| 149 | C = C.contiguous() |
| 150 | if z is not None and z.stride(-1) != 1: |
| 151 | z = z.contiguous() |
| 152 | if B.dim() == 3 and cu_seq_len is None: |
| 153 | B = B.unsqueeze(1) |
| 154 | if B.dim() == 2 and cu_seq_len is not None: |
| 155 | B = B.unsqueeze(0) |
| 156 | if C.dim() == 3 and cu_seq_len is None: |
| 157 | C = C.unsqueeze(1) |
| 158 | if C.dim() == 2 and cu_seq_len is not None: |
| 159 | C = C.unsqueeze(0) |
| 160 | |
| 161 | # Disable test_autograd_registration for now as it seems to trigger |
| 162 | # a bogus error. |
| 163 | opcheck( |
| 164 | torch.ops._C.selective_scan_fwd, |
| 165 | ( |
| 166 | u, |
| 167 | delta, |
| 168 | A, |
| 169 | B, |
| 170 | C, |
| 171 | D, |
| 172 | z, |
| 173 | delta_bias, |
| 174 | delta_softplus, |
| 175 | cu_seq_len, |
| 176 | cache_indices, |
| 177 | has_initial_state, |
| 178 | ssm_states, |
| 179 | null_block_id, |
| 180 | block_size, |
| 181 | block_idx_first_scheduled_token, |
| 182 | block_idx_last_scheduled_token, |
| 183 | initial_state_idx, |
| 184 | cu_chunk_seqlen, |
| 185 | last_chunk_indices, |
| 186 | ), |
| 187 | test_utils=["test_schema", "test_faketensor"], |
| 188 | ) |
| 189 | |
| 190 | |
| 191 | @pytest.mark.parametrize("wtype", [torch.float32]) |
| 192 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 193 | @pytest.mark.parametrize("seqlen", [128, 1024, 4096]) |
| 194 | @pytest.mark.parametrize("has_delta_bias", [True]) |
| 195 | @pytest.mark.parametrize("delta_softplus", [True]) |
| 196 | @pytest.mark.parametrize("has_z", [True]) |
| 197 | @pytest.mark.parametrize("has_D", [True]) |
| 198 | @pytest.mark.parametrize("varBC_groups", [1, 2]) |
| 199 | @pytest.mark.parametrize("is_variable_C", [True]) |
| 200 | @pytest.mark.parametrize("is_variable_B", [True]) |
| 201 | @pytest.mark.parametrize("scan_chunks", [1, 3]) |
| 202 | @skip_unless_cuda_alike |
| 203 | def test_selective_scan( |
| 204 | is_variable_B, |
| 205 | is_variable_C, |
| 206 | varBC_groups, |
| 207 | has_D, |
| 208 | has_z, |
| 209 | has_delta_bias, |
| 210 | delta_softplus, |
| 211 | seqlen, |
| 212 | itype, |
| 213 | wtype, |
| 214 | scan_chunks, |
| 215 | ): |
| 216 | if varBC_groups > 1 and (not is_variable_B or not is_variable_C): |
| 217 | pytest.skip() # This config is not applicable |
| 218 | device = "cuda" |
| 219 | rtol, atol = (6e-4, 2e-3) if itype == torch.float32 else (3e-3, 5e-3) |
| 220 | if itype == torch.bfloat16: |
| 221 | rtol, atol = 3e-2, 5e-2 |
| 222 | rtolw, atolw = (1e-3, 1e-3) |
| 223 | if has_z: # If we have z, the errors on the weights seem higher |
| 224 | rtolw = max(rtolw, rtol) |
| 225 | atolw = max(atolw, atol) |
| 226 | # set seed |
| 227 | set_random_seed(0) |
| 228 | batch_size = 1 |
| 229 | dim = 4 |
| 230 | dstate = 8 |
| 231 | A = -0.5 * torch.rand(dim, dstate, device=device, dtype=wtype) |
| 232 | A_ref = A.clone() |
| 233 | if not is_variable_B: |
| 234 | B_shape = [dim, dstate] |
| 235 | elif varBC_groups == 1: |
| 236 | B_shape = [batch_size, dstate, seqlen] |
| 237 | else: |
| 238 | B_shape = [batch_size, varBC_groups, dstate, seqlen] |
| 239 | B = torch.randn(B_shape, device=device, dtype=wtype if not is_variable_B else itype) |
| 240 | B_ref = B.clone() |
| 241 | if not is_variable_C: |
| 242 | C_shape = [dim, dstate] |
| 243 | elif varBC_groups == 1: |
| 244 | C_shape = [batch_size, dstate, seqlen] |
| 245 | else: |
| 246 | C_shape = [batch_size, varBC_groups, dstate, seqlen] |
| 247 | C = torch.randn(C_shape, device=device, dtype=wtype if not is_variable_C else itype) |
| 248 | C_ref = C.clone() |
| 249 | D = torch.randn(dim, device=device, dtype=torch.float32) if has_D else None |
| 250 | D_ref = D.clone() if D is not None else None |
| 251 | z = ( |
| 252 | torch.randn(batch_size, dim, seqlen, device=device, dtype=itype) |
| 253 | if has_z |
| 254 | else None |
| 255 | ) |
| 256 | z_ref = z.clone() if z is not None else None |
| 257 | delta_bias = ( |
| 258 | (0.5 * torch.rand(dim, device=device, dtype=torch.float32)) |
| 259 | if has_delta_bias |
| 260 | else None |
| 261 | ) |
| 262 | u = torch.randn(batch_size, dim, seqlen, device=device, dtype=itype) |
| 263 | u_ref = u.clone() |
| 264 | delta = 0.5 * torch.rand(batch_size, dim, seqlen, device=device, dtype=itype) |
| 265 | delta_ref = delta.clone() |
| 266 | state_shape = (batch_size, u.shape[1], int(A.shape[1])) |
| 267 | state = torch.randn(state_shape, device=u.device, dtype=itype, requires_grad=False) |
| 268 | state_ref = state.clone() |
| 269 | out = None |
| 270 | out_ref = None |
| 271 | outs = [] |
| 272 | for c in range(scan_chunks): |
| 273 | chunked_prompt_len = seqlen // scan_chunks |
| 274 | chunk_start = chunked_prompt_len * c |
| 275 | chunk_end = chunked_prompt_len * (c + 1) |
| 276 | if c == scan_chunks - 1: |
| 277 | chunk_end = seqlen |
| 278 | _B = B |
| 279 | if is_variable_B: |
| 280 | _B = B[..., chunk_start:chunk_end] |
| 281 | _C = C |
| 282 | if is_variable_B: |
| 283 | _C = C[..., chunk_start:chunk_end] |
| 284 | _z = z |
| 285 | if has_z: |
| 286 | assert z is not None |
| 287 | _z = z[..., chunk_start:chunk_end] |
| 288 | out = selective_scan_fn( |
| 289 | u[..., chunk_start:chunk_end], |
| 290 | state, |
| 291 | delta[..., chunk_start:chunk_end], |
| 292 | A, |
| 293 | _B, |
| 294 | _C, |
| 295 | D, |
| 296 | z=_z, |
| 297 | delta_bias=delta_bias, |
| 298 | delta_softplus=delta_softplus, |
| 299 | has_initial_state=torch.ones(batch_size, device=u.device, dtype=torch.bool) |
| 300 | if c > 0 |
| 301 | else None, |
| 302 | block_size=2048, |
| 303 | block_idx_first_scheduled_token=None, |
| 304 | block_idx_last_scheduled_token=None, |
| 305 | initial_state_idx=None, |
| 306 | ) |
| 307 | outs.append(out) |
| 308 | if len(outs) > 1: |
| 309 | out = torch.cat(outs, dim=-1) |
| 310 | |
| 311 | out_ref, state_ref, *rest = selective_scan_ref( |
| 312 | u_ref, |
| 313 | delta_ref, |
| 314 | A_ref, |
| 315 | B_ref, |
| 316 | C_ref, |
| 317 | D_ref, |
| 318 | z=z_ref, |
| 319 | delta_bias=delta_bias, |
| 320 | delta_softplus=delta_softplus, |
| 321 | return_last_state=True, |
| 322 | ) |
| 323 | |
| 324 | assert out is not None and out_ref is not None |
| 325 | assert torch.allclose(out, out_ref, rtol=rtol, atol=atol) |
| 326 | assert state is not None and state_ref is not None |
| 327 | assert torch.allclose(state, state_ref.to(itype), rtol=rtol, atol=atol) |
| 328 | |
| 329 | selective_scan_opcheck_fn( |
| 330 | u, |
| 331 | delta, |
| 332 | A, |
| 333 | B, |
| 334 | C, |
| 335 | D, |
| 336 | z, |
| 337 | delta_bias=delta_bias, |
| 338 | delta_softplus=delta_softplus, |
| 339 | ssm_states=state, |
| 340 | block_size=2048, |
| 341 | ) |
| 342 | |
| 343 | |
| 344 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 345 | @pytest.mark.parametrize("has_z", [False, True]) |
| 346 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 347 | @pytest.mark.parametrize("dim", [2048, 2048 + 16, 4096]) |
| 348 | @pytest.mark.skipif( |
| 349 | current_platform.is_cpu(), |
| 350 | reason=( |
| 351 | "CPU kernel for selective_state_update only supports " |
| 352 | "Mamba 2 (scalar A/dt), not Mamba 1." |
| 353 | ), |
| 354 | ) |
| 355 | def test_selective_state_update(dim, dstate, has_z, itype): |
| 356 | device = DEVICE |
| 357 | rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2) |
| 358 | if itype == torch.bfloat16: |
| 359 | rtol, atol = 1e-2, 5e-2 |
| 360 | if ( |
| 361 | current_platform.is_rocm() |
| 362 | or current_platform.is_xpu() |
| 363 | or current_platform.is_device_capability_family(90) |
| 364 | ): |
| 365 | atol *= 2 |
| 366 | # set seed |
| 367 | set_random_seed(0) |
| 368 | batch_size = 1 |
| 369 | state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device) |
| 370 | x = torch.randn(batch_size, dim, device=device, dtype=itype) |
| 371 | out = torch.empty_like(x) |
| 372 | dt = torch.randn(batch_size, dim, device=device, dtype=itype) |
| 373 | dt_bias = torch.rand(dim, device=device) - 4.0 |
| 374 | A = -torch.rand(dim, dstate, device=device) - 1.0 |
| 375 | B = torch.randn(batch_size, dstate, device=device) |
| 376 | C = torch.randn(batch_size, dstate, device=device) |
| 377 | D = torch.randn(dim, device=device) |
| 378 | z = torch.randn_like(x) if has_z else None |
| 379 | state_ref = state.detach().clone() |
| 380 | selective_state_update( |
| 381 | state, x, dt, A, B, C, D=D, z=z, dt_bias=dt_bias, dt_softplus=True, out=out |
| 382 | ) |
| 383 | out_ref = selective_state_update_ref( |
| 384 | state_ref, x, dt, A, B, C, D=D, z=z, dt_bias=dt_bias, dt_softplus=True |
| 385 | ) |
| 386 | |
| 387 | assert torch.allclose(state, state_ref, rtol=rtol, atol=atol) |
| 388 | assert torch.allclose(out, out_ref, rtol=rtol, atol=atol) |
| 389 | |
| 390 | |
| 391 | @pytest.mark.parametrize("philox_rounds", [0, 4]) |
| 392 | @pytest.mark.parametrize("has_z", [False, True]) |
| 393 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 394 | @pytest.mark.parametrize("dim", [2048, 4096]) |
| 395 | @pytest.mark.skipif( |
| 396 | not ( |
| 397 | current_platform.is_cuda() and current_platform.is_device_capability_family(100) |
| 398 | ), |
| 399 | reason="Stochastic rounding in triton is only supported" |
| 400 | " on compute capability 10.0 CUDA devices.", |
| 401 | ) |
| 402 | def test_selective_state_update_stochastic_rounding(dim, dstate, has_z, philox_rounds): |
| 403 | device = DEVICE |
| 404 | rtol, atol = 5e-3, 1e-1 |
| 405 | # set seed |
| 406 | set_random_seed(0) |
| 407 | batch_size = 1 |
| 408 | state = torch.randn(batch_size, dim, dstate, dtype=torch.float16, device=device) |
| 409 | x = torch.randn(batch_size, dim, device=device, dtype=torch.bfloat16) |
| 410 | out = torch.empty_like(x) |
| 411 | dt = torch.randn(batch_size, dim, device=device, dtype=torch.bfloat16) |
| 412 | dt_bias = torch.rand(dim, device=device) - 4.0 |
| 413 | A = -torch.rand(dim, dstate, device=device) - 1.0 |
| 414 | B = torch.randn(batch_size, dstate, device=device) |
| 415 | C = torch.randn(batch_size, dstate, device=device) |
| 416 | D = torch.randn(dim, device=device) |
| 417 | z = torch.randn_like(x) if has_z else None |
| 418 | # Reference uses fp32 state to get ground truth |
| 419 | state_ref = state.float() |
| 420 | selective_state_update( |
| 421 | state, |
| 422 | x, |
| 423 | dt, |
| 424 | A, |
| 425 | B, |
| 426 | C, |
| 427 | D=D, |
| 428 | z=z, |
| 429 | dt_bias=dt_bias, |
| 430 | dt_softplus=True, |
| 431 | out=out, |
| 432 | enable_stochastic_rounding=True, |
| 433 | cache_philox_rounds=philox_rounds, |
| 434 | ) |
| 435 | out_ref = selective_state_update_ref( |
| 436 | state_ref, x, dt, A, B, C, D=D, z=z, dt_bias=dt_bias, dt_softplus=True |
| 437 | ) |
| 438 | |
| 439 | assert state.dtype == torch.float16 |
| 440 | assert torch.allclose(state, state_ref.to(torch.float16), rtol=rtol, atol=atol) |
| 441 | assert torch.allclose(out, out_ref, rtol=rtol, atol=atol) |
| 442 | |
| 443 | |
| 444 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 445 | @pytest.mark.parametrize("has_z", [False, True]) |
| 446 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 447 | @pytest.mark.parametrize("dim", [2048, 2048 + 16, 4096]) |
| 448 | @pytest.mark.parametrize("max_seq_len", [1, 2, 4]) |
| 449 | @pytest.mark.skipif( |
| 450 | current_platform.is_cpu(), |
| 451 | reason=( |
| 452 | "CPU kernel for selective_state_update only supports " |
| 453 | "Mamba 2 (scalar A/dt), not Mamba 1." |
| 454 | ), |
| 455 | ) |
| 456 | def test_selective_state_update_varlen(dim, dstate, has_z, itype, max_seq_len): |
| 457 | device = DEVICE |
| 458 | rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2) |
| 459 | if itype == torch.bfloat16: |
| 460 | rtol, atol = 5e-2, 1.5e-1 |
| 461 | if current_platform.is_rocm() or current_platform.is_xpu(): |
| 462 | atol *= 2 |
| 463 | # set seed |
| 464 | set_random_seed(0) |
| 465 | batch_size = 4 |
| 466 | token_counts = torch.randint(1, max_seq_len + 1, (batch_size,), device=device) |
| 467 | total_tokens = int(token_counts.sum().item()) |
| 468 | cu_seqlens = torch.tensor( |
| 469 | [0] + torch.cumsum(token_counts, dim=0).tolist(), |
| 470 | dtype=torch.int32, |
| 471 | device=device, |
| 472 | ) |
| 473 | state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device) |
| 474 | x = torch.randn(total_tokens, dim, device=device, dtype=itype) |
| 475 | out = torch.empty_like(x) |
| 476 | dt = torch.randn(total_tokens, dim, device=device, dtype=itype) |
| 477 | dt_bias = torch.rand(dim, device=device) - 4.0 |
| 478 | A = -torch.rand(dim, dstate, device=device) - 1.0 |
| 479 | B = torch.randn(total_tokens, dstate, device=device) |
| 480 | C = torch.randn(total_tokens, dstate, device=device) |
| 481 | D = torch.randn(dim, device=device) |
| 482 | z = torch.randn_like(x) if has_z else None |
| 483 | state_ref = state.detach().clone() |
| 484 | selective_state_update( |
| 485 | state, |
| 486 | x, |
| 487 | dt, |
| 488 | A, |
| 489 | B, |
| 490 | C, |
| 491 | D=D, |
| 492 | z=z, |
| 493 | dt_bias=dt_bias, |
| 494 | dt_softplus=True, |
| 495 | out=out, |
| 496 | cu_seqlens=cu_seqlens, |
| 497 | ) |
| 498 | |
| 499 | out_ref_list = [] |
| 500 | for seq_idx in range(batch_size): |
| 501 | start_idx = cu_seqlens[seq_idx].item() |
| 502 | end_idx = cu_seqlens[seq_idx + 1].item() |
| 503 | num_tokens = end_idx - start_idx |
| 504 | for token_idx in range(num_tokens): |
| 505 | idx = start_idx + token_idx |
| 506 | out_ref_list.append( |
| 507 | selective_state_update_ref( |
| 508 | state_ref[seq_idx : seq_idx + 1], |
| 509 | x[idx : idx + 1], |
| 510 | dt[idx : idx + 1], |
| 511 | A, |
| 512 | B[idx : idx + 1], |
| 513 | C[idx : idx + 1], |
| 514 | D=D, |
| 515 | z=z[idx : idx + 1] if z is not None else None, |
| 516 | dt_bias=dt_bias, |
| 517 | dt_softplus=True, |
| 518 | ) |
| 519 | ) |
| 520 | out_ref = torch.cat(out_ref_list, dim=0) |
| 521 | assert torch.allclose(state, state_ref, rtol=rtol, atol=atol) |
| 522 | assert torch.allclose(out, out_ref, rtol=rtol, atol=atol) |
| 523 | |
| 524 | |
| 525 | @pytest.mark.parametrize("wtype", [torch.float32]) |
| 526 | @pytest.mark.parametrize("itype", [torch.float32]) |
| 527 | @pytest.mark.parametrize("seqlen", [1, 256, 1024, 4096]) |
| 528 | @pytest.mark.parametrize("return_last_state", [True]) |
| 529 | @pytest.mark.parametrize("has_delta_bias", [True]) |
| 530 | @pytest.mark.parametrize("delta_softplus", [True]) |
| 531 | @pytest.mark.parametrize("has_z", [True]) |
| 532 | @pytest.mark.parametrize("has_D", [True]) |
| 533 | @pytest.mark.parametrize("varBC_groups", [1, 2]) |
| 534 | @pytest.mark.parametrize("is_variable_C", [True]) |
| 535 | @pytest.mark.parametrize("is_variable_B", [True]) |
| 536 | # tests correctness in case subset of the sequences are padded |
| 537 | @pytest.mark.parametrize("with_padding", [False, True]) |
| 538 | @skip_unless_cuda_alike |
| 539 | def test_selective_scan_varlen( |
| 540 | with_padding, |
| 541 | is_variable_B, |
| 542 | is_variable_C, |
| 543 | varBC_groups, |
| 544 | has_D, |
| 545 | has_z, |
| 546 | has_delta_bias, |
| 547 | delta_softplus, |
| 548 | return_last_state, |
| 549 | seqlen, |
| 550 | itype, |
| 551 | wtype, |
| 552 | ): |
| 553 | if varBC_groups > 1 and (not is_variable_B or not is_variable_C): |
| 554 | pytest.skip() # This config is not applicable |
| 555 | device = "cuda" |
| 556 | rtol, atol = (6e-4, 2e-3) if itype == torch.float32 else (3e-3, 5e-3) |
| 557 | if itype == torch.bfloat16: |
| 558 | rtol, atol = 3e-2, 5e-2 |
| 559 | rtolw, atolw = (1e-3, 1e-3) |
| 560 | if has_z: # If we have z, the errors on the weights seem higher |
| 561 | rtolw = max(rtolw, rtol) |
| 562 | atolw = max(atolw, atol) |
| 563 | # set seed |
| 564 | torch.random.manual_seed(0) |
| 565 | seqlens = [] |
| 566 | batch_size = 4 |
| 567 | if seqlen < 10: |
| 568 | batch_size = 1 |
| 569 | padding = 3 if with_padding else 0 |
| 570 | padded_batch_size = batch_size + padding |
| 571 | |
| 572 | if with_padding and seqlen < padded_batch_size: |
| 573 | pytest.skip() |
| 574 | |
| 575 | nsplits = padded_batch_size - 1 |
| 576 | eos_pos = torch.randperm(seqlen - 1)[:nsplits].sort().values |
| 577 | seqlens.append( |
| 578 | torch.diff( |
| 579 | torch.cat([torch.tensor([-1]), eos_pos, torch.tensor([seqlen - 1])]) |
| 580 | ).tolist() |
| 581 | ) |
| 582 | |
| 583 | assert sum(seqlens[-1]) == seqlen |
| 584 | assert all(s > 0 for s in seqlens[-1]) |
| 585 | |
| 586 | total_entries = batch_size * 10 |
| 587 | cumsum = torch.cumsum(torch.tensor(seqlens[0]), dim=0).to(torch.int32) |
| 588 | cumsum = torch.concat([torch.tensor([0], dtype=torch.int32), cumsum], dim=0).cuda() |
| 589 | |
| 590 | dim = 4 |
| 591 | dstate = 8 |
| 592 | A = -0.5 * torch.rand(dim, dstate, device=device, dtype=wtype) |
| 593 | A_ref = A.clone() |
| 594 | B_shape = [varBC_groups, dstate, seqlen] |
| 595 | B = torch.randn(B_shape, device=device, dtype=wtype if not is_variable_B else itype) |
| 596 | B_ref = B.clone() |
| 597 | C_shape = [varBC_groups, dstate, seqlen] |
| 598 | C = torch.randn(C_shape, device=device, dtype=wtype if not is_variable_C else itype) |
| 599 | C_ref = C.clone() |
| 600 | D = torch.randn(dim, device=device, dtype=torch.float32) if has_D else None |
| 601 | D_ref = D.clone() if D is not None else None |
| 602 | z = torch.randn(dim, seqlen, device=device, dtype=itype) |
| 603 | z_ref = z.clone() |
| 604 | delta_bias = ( |
| 605 | (0.5 * torch.rand(dim, device=device, dtype=torch.float32)) |
| 606 | if has_delta_bias |
| 607 | else None |
| 608 | ) |
| 609 | u = torch.randn(dim, seqlen, device=device, dtype=itype) |
| 610 | u_ref = u.clone() |
| 611 | delta = 0.5 * torch.rand(dim, seqlen, device=device, dtype=itype) |
| 612 | delta_ref = delta.clone() |
| 613 | out = None |
| 614 | out_ref = None |
| 615 | |
| 616 | prev_state_shape = (total_entries, u.shape[0], int(A.shape[1])) |
| 617 | prev_state = torch.randn( |
| 618 | prev_state_shape, device=u.device, dtype=itype, requires_grad=False |
| 619 | ) |
| 620 | prev_state_ref = prev_state.clone() |
| 621 | # +1 to exclude index 0 (null block) |
| 622 | state_indices = ( |
| 623 | torch.randperm(total_entries - 1, dtype=torch.int32, device=u.device)[ |
| 624 | :batch_size |
| 625 | ] |
| 626 | + 1 |
| 627 | ) |
| 628 | unused_states_bool = torch.ones(total_entries, dtype=torch.bool, device=device) |
| 629 | unused_states_bool[state_indices] = False |
| 630 | padded_state_indices = torch.concat( |
| 631 | [ |
| 632 | state_indices, |
| 633 | torch.as_tensor( |
| 634 | [NULL_BLOCK_ID] * padding, dtype=torch.int32, device=device |
| 635 | ), |
| 636 | ], |
| 637 | dim=-1, |
| 638 | ) |
| 639 | |
| 640 | has_initial_state = torch.randint( |
| 641 | 0, 2, (cumsum.shape[0] - 1,), dtype=torch.bool, device=u.device |
| 642 | ) |
| 643 | out = selective_scan_fn( |
| 644 | u, |
| 645 | prev_state, |
| 646 | delta, |
| 647 | A, |
| 648 | B, |
| 649 | C, |
| 650 | D, |
| 651 | z, |
| 652 | delta_bias, |
| 653 | delta_softplus, |
| 654 | cumsum, |
| 655 | padded_state_indices, |
| 656 | has_initial_state, |
| 657 | ) |
| 658 | outs_ref = [] |
| 659 | splits = [ |
| 660 | torch.split(var, seqlens[0], dim=-1) |
| 661 | for var in (u_ref, delta_ref, B_ref, C_ref, z_ref) |
| 662 | ] |
| 663 | for i in range(len(seqlens[0])): |
| 664 | u_s, delta_s, B_s, C_s, z_s = (v[i].unsqueeze(0) for v in splits) |
| 665 | if padded_state_indices[i] == NULL_BLOCK_ID: |
| 666 | continue |
| 667 | out_ref_s, _ = selective_scan_ref( |
| 668 | u_s, |
| 669 | delta_s, |
| 670 | A_ref, |
| 671 | B_s, |
| 672 | C_s, |
| 673 | D_ref, |
| 674 | z=z_s, |
| 675 | delta_bias=delta_bias, |
| 676 | delta_softplus=delta_softplus, |
| 677 | return_last_state=return_last_state, |
| 678 | prev_state=prev_state_ref[padded_state_indices[i]].unsqueeze(0) |
| 679 | if has_initial_state[i] |
| 680 | else None, |
| 681 | final_state_out=prev_state_ref[padded_state_indices[i]].unsqueeze(0), |
| 682 | ) |
| 683 | outs_ref.append(out_ref_s) |
| 684 | out_ref = torch.cat(outs_ref, dim=-1)[0] |
| 685 | |
| 686 | unpadded_out = out[:, : out_ref[0].shape[-1]] |
| 687 | print("Output diff max", (unpadded_out - out_ref).max()) |
| 688 | print("Output diff mean", (unpadded_out - out_ref).mean()) |
| 689 | print("Output state diff max", (prev_state - prev_state_ref).max()) |
| 690 | print("Output state diff mean", (prev_state - prev_state_ref).mean()) |
| 691 | assert torch.allclose(prev_state, prev_state_ref, rtol=rtol, atol=atol) |
| 692 | assert torch.allclose(unpadded_out, out_ref, rtol=rtol, atol=atol) |
| 693 | selective_scan_opcheck_fn( |
| 694 | u, |
| 695 | delta, |
| 696 | A, |
| 697 | B, |
| 698 | C, |
| 699 | D, |
| 700 | z, |
| 701 | delta_bias, |
| 702 | delta_softplus, |
| 703 | cumsum, |
| 704 | padded_state_indices, |
| 705 | has_initial_state, |
| 706 | prev_state, |
| 707 | block_size=2048, |
| 708 | ) |
| 709 | |
| 710 | |
| 711 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 712 | @pytest.mark.parametrize("has_z", [True]) |
| 713 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 714 | @pytest.mark.parametrize("dim", [2048, 2048 + 16, 4096]) |
| 715 | # tests correctness in case subset of the sequences are padded |
| 716 | @pytest.mark.parametrize("with_padding", [True, False]) |
| 717 | @pytest.mark.skipif( |
| 718 | current_platform.is_cpu(), |
| 719 | reason=( |
| 720 | "CPU kernel for selective_state_update only supports " |
| 721 | "Mamba 2 (scalar A/dt), not Mamba 1." |
| 722 | ), |
| 723 | ) |
| 724 | def test_selective_state_update_with_batch_indices( |
| 725 | with_padding, dim, dstate, has_z, itype |
| 726 | ): |
| 727 | device = DEVICE |
| 728 | rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2) |
| 729 | if itype == torch.bfloat16: |
| 730 | rtol, atol = 1e-1, 1e-1 |
| 731 | if current_platform.is_rocm() or current_platform.is_xpu(): |
| 732 | atol *= 2 |
| 733 | # set seed |
| 734 | torch.random.manual_seed(0) |
| 735 | batch_size = 3 |
| 736 | padding = 5 if with_padding else 0 |
| 737 | padded_batch_size = batch_size + padding |
| 738 | total_entries = 10 * batch_size |
| 739 | state = torch.randn(total_entries, dim, dstate, dtype=itype, device=device) |
| 740 | # +1 to exclude index 0 (null block) |
| 741 | state_indices = (torch.randperm(total_entries - 1)[:batch_size] + 1).to( |
| 742 | dtype=torch.int32, device=device |
| 743 | ) |
| 744 | unused_states_bool = torch.ones(total_entries, dtype=torch.bool, device=device) |
| 745 | unused_states_bool[state_indices] = False |
| 746 | padded_state_indices = torch.concat( |
| 747 | [ |
| 748 | state_indices, |
| 749 | torch.as_tensor( |
| 750 | [NULL_BLOCK_ID] * padding, dtype=torch.int32, device=device |
| 751 | ), |
| 752 | ], |
| 753 | dim=0, |
| 754 | ) |
| 755 | x = torch.randn(padded_batch_size, dim, device=device, dtype=itype) |
| 756 | out = torch.empty_like(x) |
| 757 | dt = torch.randn(padded_batch_size, dim, device=device, dtype=itype) |
| 758 | dt_bias = torch.rand(dim, device=device) - 4.0 |
| 759 | A = -torch.rand(dim, dstate, device=device) - 1.0 |
| 760 | B = torch.randn(padded_batch_size, dstate, device=device) |
| 761 | C = torch.randn(padded_batch_size, dstate, device=device) |
| 762 | D = torch.randn(dim, device=device) |
| 763 | z = torch.randn_like(x) if has_z else None |
| 764 | state_ref = state[state_indices, :].clone() |
| 765 | state_before = state.clone() |
| 766 | selective_state_update( |
| 767 | state, |
| 768 | x, |
| 769 | dt, |
| 770 | A, |
| 771 | B, |
| 772 | C, |
| 773 | D=D, |
| 774 | z=z, |
| 775 | dt_bias=dt_bias, |
| 776 | dt_softplus=True, |
| 777 | state_batch_indices=padded_state_indices, |
| 778 | out=out, |
| 779 | ) |
| 780 | out_ref = selective_state_update_ref( |
| 781 | state_ref, |
| 782 | x[:batch_size], |
| 783 | dt[:batch_size], |
| 784 | A, |
| 785 | B[:batch_size], |
| 786 | C[:batch_size], |
| 787 | D=D, |
| 788 | z=z[:batch_size] if z is not None else None, |
| 789 | dt_bias=dt_bias, |
| 790 | dt_softplus=True, |
| 791 | ) |
| 792 | |
| 793 | print("Output diff max", (out[:batch_size] - out_ref).max()) |
| 794 | print("Output diff mean", (out[:batch_size] - out_ref).mean()) |
| 795 | print("Output state diff max", (state[state_indices, :] - state_ref).max()) |
| 796 | print("Output state diff mean", (state[state_indices, :] - state_ref).mean()) |
| 797 | # test padded entries stay the same |
| 798 | if with_padding: |
| 799 | assert torch.equal(state_before[unused_states_bool], state[unused_states_bool]) |
| 800 | assert torch.equal(x[batch_size + 1 :], x[batch_size + 1 :]) |
| 801 | assert torch.equal(dt[batch_size + 1 :], dt[batch_size + 1 :]) |
| 802 | assert torch.equal(B[batch_size + 1 :], B[batch_size + 1 :]) |
| 803 | assert torch.equal(C[batch_size + 1 :], C[batch_size + 1 :]) |
| 804 | |
| 805 | # test "real" entries |
| 806 | assert torch.allclose(state[state_indices, :], state_ref, rtol=rtol, atol=atol) |
| 807 | assert torch.allclose(out[:batch_size], out_ref, rtol=rtol, atol=atol) |
| 808 | |
| 809 | |
| 810 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 811 | @pytest.mark.parametrize("has_z", [False, True]) |
| 812 | @pytest.mark.parametrize("tie_hdim", [False, True]) |
| 813 | @pytest.mark.parametrize("ngroups", [1, 4]) |
| 814 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 815 | @pytest.mark.parametrize("dim", [2048, 4096]) |
| 816 | @pytest.mark.skipif( |
| 817 | current_platform.is_cpu(), |
| 818 | reason=( |
| 819 | "CPU kernel for selective_state_update only supports " |
| 820 | "Mamba 2 (scalar A/dt), not Mamba 1." |
| 821 | ), |
| 822 | ) |
| 823 | def test_selective_state_update_with_heads_with_batch_indices( |
| 824 | dim, dstate, ngroups, has_z, tie_hdim, itype |
| 825 | ): |
| 826 | device = DEVICE |
| 827 | rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 3e-2) |
| 828 | if itype == torch.bfloat16: |
| 829 | rtol, atol = 1e-1, 1e-1 |
| 830 | # set seed |
| 831 | torch.random.manual_seed(0) |
| 832 | batch_size = 3 |
| 833 | headdim = 64 |
| 834 | nheads = dim // headdim |
| 835 | |
| 836 | total_entries = 10 * batch_size |
| 837 | state = torch.randn( |
| 838 | total_entries, nheads, headdim, dstate, dtype=itype, device=device |
| 839 | ) |
| 840 | # +1 to exclude index 0 (null block) |
| 841 | state_indices = (torch.randperm(total_entries - 1)[:batch_size] + 1).to( |
| 842 | dtype=torch.int32, device=device |
| 843 | ) |
| 844 | |
| 845 | x = torch.randn(batch_size, nheads, headdim, device=device, dtype=itype) |
| 846 | out = torch.empty_like(x) |
| 847 | if not tie_hdim: |
| 848 | dt = torch.randn(batch_size, nheads, headdim, device=device, dtype=itype) |
| 849 | dt_bias = torch.rand(nheads, headdim, device=device) - 4.0 |
| 850 | A = -torch.rand(nheads, headdim, dstate, device=device) - 1.0 |
| 851 | D = torch.randn(nheads, headdim, device=device) |
| 852 | else: |
| 853 | dt = repeat( |
| 854 | torch.randn(batch_size, nheads, device=device, dtype=itype), |
| 855 | "b h -> b h p", |
| 856 | p=headdim, |
| 857 | ) |
| 858 | dt_bias = repeat(torch.rand(nheads, device=device) - 4.0, "h -> h p", p=headdim) |
| 859 | A = repeat( |
| 860 | -torch.rand(nheads, device=device) - 1.0, "h -> h p n", p=headdim, n=dstate |
| 861 | ) |
| 862 | D = repeat(torch.randn(nheads, device=device), "h -> h p", p=headdim) |
| 863 | B = torch.randn(batch_size, ngroups, dstate, device=device) |
| 864 | C = torch.randn(batch_size, ngroups, dstate, device=device) |
| 865 | z = torch.randn_like(x) if has_z else None |
| 866 | state_ref = state[state_indices, :].detach().clone() |
| 867 | selective_state_update( |
| 868 | state, |
| 869 | x, |
| 870 | dt, |
| 871 | A, |
| 872 | B, |
| 873 | C, |
| 874 | D=D, |
| 875 | z=z, |
| 876 | dt_bias=dt_bias, |
| 877 | dt_softplus=True, |
| 878 | state_batch_indices=state_indices, |
| 879 | out=out, |
| 880 | ) |
| 881 | out_ref = selective_state_update_ref( |
| 882 | state_ref, x, dt, A, B, C, D=D, z=z, dt_bias=dt_bias, dt_softplus=True |
| 883 | ) |
| 884 | |
| 885 | print(f"Output max diff: {(out - out_ref).abs().max().item()}") |
| 886 | print(f"Output mean diff: {(out - out_ref).abs().mean().item()}") |
| 887 | assert torch.allclose(state[state_indices, :], state_ref, rtol=rtol, atol=atol) |
| 888 | assert torch.allclose(out, out_ref, rtol=rtol, atol=atol) |
| 889 | |
| 890 | |
| 891 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 892 | @pytest.mark.parametrize("has_z", [False, True]) |
| 893 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 894 | @pytest.mark.parametrize("dim", [2048, 4096]) |
| 895 | @pytest.mark.parametrize("max_seq_len", [2, 4]) |
| 896 | @pytest.mark.skipif( |
| 897 | current_platform.is_cpu(), |
| 898 | reason=( |
| 899 | "CPU kernel for selective_state_update only supports " |
| 900 | "Mamba 2 (scalar A/dt), not Mamba 1." |
| 901 | ), |
| 902 | ) |
| 903 | def test_selective_state_update_with_num_accepted_tokens( |
| 904 | dim, dstate, has_z, itype, max_seq_len |
| 905 | ): |
| 906 | device = DEVICE |
| 907 | rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2) |
| 908 | if itype == torch.bfloat16: |
| 909 | rtol, atol = 5e-2, 1.5e-1 |
| 910 | if current_platform.is_rocm() or current_platform.is_xpu(): |
| 911 | atol *= 2 |
| 912 | |
| 913 | set_random_seed(0) |
| 914 | batch_size = 4 |
| 915 | |
| 916 | tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device) |
| 917 | total_tokens = int(tokens_per_seq.sum().item()) |
| 918 | |
| 919 | num_accepted_tokens = torch.randint(0, max_seq_len, (batch_size,), device=device) |
| 920 | num_accepted_tokens[0] = 0 # Add edge-case of no accepted tokens |
| 921 | num_accepted_tokens[1] = max_seq_len # Add edge-case of all tokens accepted |
| 922 | |
| 923 | cu_seqlens = torch.tensor( |
| 924 | [0] + torch.cumsum(tokens_per_seq, dim=0).tolist(), |
| 925 | dtype=torch.int32, |
| 926 | device=device, |
| 927 | ) |
| 928 | |
| 929 | total_state_slots = 50 |
| 930 | state = torch.randn(total_state_slots, dim, dstate, dtype=itype, device=device) |
| 931 | |
| 932 | state_batch_indices = torch.full( |
| 933 | (batch_size, max_seq_len), NULL_BLOCK_ID, dtype=torch.int32, device=device |
| 934 | ) |
| 935 | # Start from 1 to exclude null block at index 0 |
| 936 | initial_state_slots = torch.randint( |
| 937 | 1, 15, (batch_size,), device=device, dtype=torch.int32 |
| 938 | ) |
| 939 | for seq_idx in range(batch_size): |
| 940 | token_pos = max(num_accepted_tokens[seq_idx].item() - 1, 0) |
| 941 | state_batch_indices[seq_idx, token_pos] = initial_state_slots[seq_idx] |
| 942 | |
| 943 | dst_state_batch_indices = torch.full( |
| 944 | (batch_size, max_seq_len), NULL_BLOCK_ID, dtype=torch.int32, device=device |
| 945 | ) |
| 946 | slot_offset = 15 |
| 947 | dst_slots_map = {} |
| 948 | for seq_idx in range(batch_size): |
| 949 | for token_idx in range(tokens_per_seq[seq_idx].item()): |
| 950 | dst_state_batch_indices[seq_idx, token_idx] = slot_offset |
| 951 | dst_slots_map[(seq_idx, token_idx)] = slot_offset |
| 952 | slot_offset += 1 |
| 953 | |
| 954 | x = torch.randn(total_tokens, dim, device=device, dtype=itype) |
| 955 | out = torch.empty_like(x) |
| 956 | dt = torch.randn(total_tokens, dim, device=device, dtype=itype) |
| 957 | dt_bias = torch.rand(dim, device=device) - 4.0 |
| 958 | A = -torch.rand(dim, dstate, device=device) - 1.0 |
| 959 | B = torch.randn(total_tokens, dstate, device=device) |
| 960 | C = torch.randn(total_tokens, dstate, device=device) |
| 961 | D = torch.randn(dim, device=device) |
| 962 | z = torch.randn_like(x) if has_z else None |
| 963 | |
| 964 | state_ref_intermediate = {} |
| 965 | out_ref_list = [] |
| 966 | |
| 967 | for seq_idx in range(batch_size): |
| 968 | seq_start = cu_seqlens[seq_idx].item() |
| 969 | seq_end = cu_seqlens[seq_idx + 1].item() |
| 970 | num_tokens = seq_end - seq_start |
| 971 | |
| 972 | token_pos = max(num_accepted_tokens[seq_idx].item() - 1, 0) |
| 973 | initial_slot = state_batch_indices[seq_idx, token_pos].item() |
| 974 | state_seq = state[initial_slot : initial_slot + 1].clone() |
| 975 | |
| 976 | for token_idx in range(num_tokens): |
| 977 | global_idx = seq_start + token_idx |
| 978 | |
| 979 | out_token = selective_state_update_ref( |
| 980 | state_seq, |
| 981 | x[global_idx : global_idx + 1], |
| 982 | dt[global_idx : global_idx + 1], |
| 983 | A, |
| 984 | B[global_idx : global_idx + 1], |
| 985 | C[global_idx : global_idx + 1], |
| 986 | D=D, |
| 987 | z=z[global_idx : global_idx + 1] if z is not None else None, |
| 988 | dt_bias=dt_bias, |
| 989 | dt_softplus=True, |
| 990 | ) |
| 991 | out_ref_list.append(out_token) |
| 992 | state_ref_intermediate[(seq_idx, token_idx)] = state_seq.clone() |
| 993 | |
| 994 | out_ref = torch.cat(out_ref_list, dim=0) |
| 995 | |
| 996 | selective_state_update( |
| 997 | state, |
| 998 | x, |
| 999 | dt, |
| 1000 | A, |
| 1001 | B, |
| 1002 | C, |
| 1003 | D=D, |
| 1004 | z=z, |
| 1005 | dt_bias=dt_bias, |
| 1006 | dt_softplus=True, |
| 1007 | out=out, |
| 1008 | cu_seqlens=cu_seqlens, |
| 1009 | state_batch_indices=state_batch_indices, |
| 1010 | dst_state_batch_indices=dst_state_batch_indices, |
| 1011 | num_accepted_tokens=num_accepted_tokens, |
| 1012 | ) |
| 1013 | |
| 1014 | assert torch.allclose(out, out_ref, rtol=rtol, atol=atol) |
| 1015 | |
| 1016 | for seq_idx in range(batch_size): |
| 1017 | num_tokens = tokens_per_seq[seq_idx].item() |
| 1018 | for token_idx in range(num_tokens): |
| 1019 | dst_slot = dst_slots_map[(seq_idx, token_idx)] |
| 1020 | state_ref = state_ref_intermediate[(seq_idx, token_idx)].squeeze(0) |
| 1021 | assert torch.allclose(state[dst_slot], state_ref, rtol=rtol, atol=atol) |
| 1022 | |
| 1023 | |
| 1024 | @pytest.mark.parametrize("itype", [torch.float32, torch.bfloat16]) |
| 1025 | @pytest.mark.parametrize("has_z", [False, True]) |
| 1026 | @pytest.mark.parametrize("dstate", [16, 64]) |
| 1027 | @pytest.mark.parametrize("dim", [2048, 4096]) |
| 1028 | @pytest.mark.parametrize("max_seq_len", [2, 4]) |
| 1029 | @pytest.mark.skipif( |
| 1030 | current_platform.is_cpu(), |
| 1031 | reason=( |
| 1032 | "CPU kernel for selective_state_update only supports " |
| 1033 | "Mamba 2 (scalar A/dt), not Mamba 1." |
| 1034 | ), |
| 1035 | ) |
| 1036 | def test_selective_state_update_varlen_with_num_accepted( |
| 1037 | dim, dstate, has_z, itype, max_seq_len |
| 1038 | ): |
| 1039 | device = DEVICE |
| 1040 | rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2) |
| 1041 | if itype == torch.bfloat16: |
| 1042 | rtol, atol = 5e-2, 1.5e-1 |
| 1043 | if current_platform.is_rocm() or current_platform.is_xpu(): |
| 1044 | atol *= 2 |
| 1045 | |
| 1046 | set_random_seed(0) |
| 1047 | batch_size = 4 |
| 1048 | |
| 1049 | tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device) |
| 1050 | total_tokens = int(tokens_per_seq.sum().item()) |
| 1051 | |
| 1052 | num_accepted_tokens = torch.randint(0, max_seq_len, (batch_size,), device=device) |
| 1053 | num_accepted_tokens[0] = 0 # Add edge-case of no accepted tokens |
| 1054 | num_accepted_tokens[1] = max_seq_len # Add edge-case of all tokens accepted |
| 1055 | |
| 1056 | cu_seqlens = torch.tensor( |
| 1057 | [0] + torch.cumsum(tokens_per_seq, dim=0).tolist(), |
| 1058 | dtype=torch.int32, |
| 1059 | device=device, |
| 1060 | ) |
| 1061 | |
| 1062 | total_state_slots = 50 |
| 1063 | state = torch.randn(total_state_slots, dim, dstate, dtype=itype, device=device) |
| 1064 | |
| 1065 | state_batch_indices = torch.full( |
| 1066 | (batch_size, max_seq_len), NULL_BLOCK_ID, dtype=torch.int32, device=device |
| 1067 | ) |
| 1068 | |
| 1069 | # Start from 1 to exclude null block at index 0 |
| 1070 | initial_state_slots = torch.randint( |
| 1071 | 1, 15, (batch_size,), device=device, dtype=torch.int32 |
| 1072 | ) |
| 1073 | for seq_idx in range(batch_size): |
| 1074 | token_pos = max(num_accepted_tokens[seq_idx].item() - 1, 0) |
| 1075 | state_batch_indices[seq_idx, token_pos] = initial_state_slots[seq_idx] |
| 1076 | |
| 1077 | dst_state_batch_indices = torch.full( |
| 1078 | (batch_size, max_seq_len), NULL_BLOCK_ID, dtype=torch.int32, device=device |
| 1079 | ) |
| 1080 | |
| 1081 | slot_offset = 15 |
| 1082 | dst_slots_map = {} |
| 1083 | for seq_idx in range(batch_size): |
| 1084 | for token_idx in range(tokens_per_seq[seq_idx].item()): |
| 1085 | dst_state_batch_indices[seq_idx, token_idx] = slot_offset |
| 1086 | dst_slots_map[(seq_idx, token_idx)] = slot_offset |
| 1087 | slot_offset += 1 |
| 1088 | |
| 1089 | x = torch.randn(total_tokens, dim, device=device, dtype=itype) |
| 1090 | out = torch.empty_like(x) |
| 1091 | dt = torch.randn(total_tokens, dim, device=device, dtype=itype) |
| 1092 | dt_bias = torch.rand(dim, device=device) - 4.0 |
| 1093 | A = -torch.rand(dim, dstate, device=device) - 1.0 |
| 1094 | B = torch.randn(total_tokens, dstate, device=device) |
| 1095 | C = torch.randn(total_tokens, dstate, device=device) |
| 1096 | D = torch.randn(dim, device=device) |
| 1097 | z = torch.randn_like(x) if has_z else None |
| 1098 | |
| 1099 | state_ref_intermediate = {} |
| 1100 | |
| 1101 | for seq_idx in range(batch_size): |
| 1102 | seq_start = cu_seqlens[seq_idx].item() |
| 1103 | seq_end = cu_seqlens[seq_idx + 1].item() |
| 1104 | num_tokens = seq_end - seq_start |
| 1105 | |
| 1106 | token_pos = max(num_accepted_tokens[seq_idx].item() - 1, 0) |
| 1107 | initial_slot = state_batch_indices[seq_idx, token_pos].item() |
| 1108 | state_seq = state[initial_slot : initial_slot + 1].clone() |
| 1109 | |
| 1110 | for token_idx in range(num_tokens): |
| 1111 | global_idx = seq_start + token_idx |
| 1112 | |
| 1113 | selective_state_update_ref( |
| 1114 | state_seq, |
| 1115 | x[global_idx : global_idx + 1], |
| 1116 | dt[global_idx : global_idx + 1], |
| 1117 | A, |
| 1118 | B[global_idx : global_idx + 1], |
| 1119 | C[global_idx : global_idx + 1], |
| 1120 | D=D, |
| 1121 | z=z[global_idx : global_idx + 1] if z is not None else None, |
| 1122 | dt_bias=dt_bias, |
| 1123 | dt_softplus=True, |
| 1124 | ) |
| 1125 | |
| 1126 | state_ref_intermediate[(seq_idx, token_idx)] = state_seq.clone() |
| 1127 | |
| 1128 | selective_state_update( |
| 1129 | state, |
| 1130 | x, |
| 1131 | dt, |
| 1132 | A, |
| 1133 | B, |
| 1134 | C, |
| 1135 | D=D, |
| 1136 | z=z, |
| 1137 | dt_bias=dt_bias, |
| 1138 | dt_softplus=True, |
| 1139 | out=out, |
| 1140 | cu_seqlens=cu_seqlens, |
| 1141 | state_batch_indices=state_batch_indices, |
| 1142 | dst_state_batch_indices=dst_state_batch_indices, |
| 1143 | num_accepted_tokens=num_accepted_tokens, |
| 1144 | ) |
| 1145 | |
| 1146 | for seq_idx in range(batch_size): |
| 1147 | num_tokens = tokens_per_seq[seq_idx].item() |
| 1148 | |
| 1149 | for token_idx in range(num_tokens): |
| 1150 | dst_slot = dst_slots_map[(seq_idx, token_idx)] |
| 1151 | state_ref = state_ref_intermediate[(seq_idx, token_idx)].squeeze(0) |
| 1152 | |
| 1153 | assert torch.allclose(state[dst_slot], state_ref, rtol=rtol, atol=atol) |