| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """ROCm-specific tests for FP8 quantization and kernels.""" |
| 4 | |
| 5 | import importlib |
| 6 | import math |
| 7 | import warnings |
| 8 | |
| 9 | import pytest |
| 10 | import torch |
| 11 | |
| 12 | import vllm._custom_ops as ops |
| 13 | from tests.kernels.quant_utils import ( |
| 14 | ref_dynamic_per_tensor_fp8_quant, |
| 15 | ref_dynamic_per_token_quant, |
| 16 | ) |
| 17 | from tests.kernels.utils import _assert_deterministic |
| 18 | from vllm.platforms import current_platform |
| 19 | from vllm.utils.torch_utils import set_random_seed |
| 20 | |
| 21 | pytestmark = pytest.mark.skipif( |
| 22 | not current_platform.is_rocm(), reason="ROCm-specific tests" |
| 23 | ) |
| 24 | |
| 25 | DEVICE = "cuda" |
| 26 | FP8_DTYPE = current_platform.fp8_dtype() |
| 27 | fp8_only = pytest.mark.skipif( |
| 28 | not current_platform.supports_fp8(), reason="ROCm FP8 only" |
| 29 | ) |
| 30 | |
| 31 | ENV_DEFAULTS = { |
| 32 | "VLLM_ROCM_FP8_PADDING": True, |
| 33 | "VLLM_ROCM_USE_AITER_FP8BMM": True, |
| 34 | "VLLM_ROCM_FP8_MFMA_PAGE_ATTN": False, |
| 35 | } |
| 36 | |
| 37 | SCALED_MM_CASES = [ |
| 38 | pytest.param((1, 128, 128), 0.05, 0.15, 0.95, id="1x128x128"), |
| 39 | pytest.param((4, 256, 512), 0.10, 0.30, 0.96, id="4x256x512"), |
| 40 | pytest.param((8, 512, 1024), 0.20, 0.50, 0.99, id="8x512x1024"), |
| 41 | ] |
| 42 | |
| 43 | |
| 44 | def _reload_envs(): |
| 45 | import vllm.envs as envs |
| 46 | |
| 47 | return importlib.reload(envs) |
| 48 | |
| 49 | |
| 50 | @pytest.fixture(autouse=True) |
| 51 | def _restore_rocm_env_state(): |
| 52 | """Restore global env + AITER flag state after every test. |
| 53 | |
| 54 | Tests here reload ``vllm.envs`` (a module singleton) and refresh the |
| 55 | cached ``rocm_aiter_ops`` env flags. Restoring in a fixture teardown |
| 56 | guarantees cleanup even when an assertion fails mid-test, preventing |
| 57 | state from leaking into other tests in the same worker. |
| 58 | """ |
| 59 | yield |
| 60 | _reload_envs() |
| 61 | from vllm._aiter_ops import rocm_aiter_ops |
| 62 | |
| 63 | rocm_aiter_ops.refresh_env_variables() |
| 64 | |
| 65 | |
| 66 | def _quantile(flat: torch.Tensor, q: float) -> float: |
| 67 | if flat.numel() == 1: |
| 68 | return flat.item() |
| 69 | return torch.quantile(flat, q).item() |
| 70 | |
| 71 | |
| 72 | def _format_observed_rate(count: int, total: int) -> str: |
| 73 | return f"{count / total:.4%} ({count}/{total})" |
| 74 | |
| 75 | |
| 76 | def _format_allowed_rate(rate: float, total: int) -> str: |
| 77 | allowed_count = math.floor(rate * total) |
| 78 | return f"{rate:.4%} (<= {allowed_count}/{total})" |
| 79 | |
| 80 | |
| 81 | def _assert_abs_error_budget( |
| 82 | actual: torch.Tensor, |
| 83 | expected: torch.Tensor, |
| 84 | *, |
| 85 | label: str, |
| 86 | tight_atol: float, |
| 87 | max_atol: float, |
| 88 | pass_rate: float, |
| 89 | ) -> None: |
| 90 | diff = (actual.float() - expected.float()).abs().flatten() |
| 91 | total = diff.numel() |
| 92 | within_tight_count = int((diff <= tight_atol).sum().item()) |
| 93 | fail_count = total - within_tight_count |
| 94 | above_max_count = int((diff > max_atol).sum().item()) |
| 95 | within_tight = within_tight_count / total |
| 96 | allowed_fail_rate = 1.0 - pass_rate |
| 97 | above_max = above_max_count / total |
| 98 | max_fail_rate = 0.0 |
| 99 | max_diff = diff.max().item() |
| 100 | mean_diff = diff.mean().item() |
| 101 | p99 = _quantile(diff, 0.99) |
| 102 | p999 = _quantile(diff, 0.999) |
| 103 | msg = ( |
| 104 | f"[rocm_fp8] {label}: abs<={tight_atol} pass={within_tight:.4%} " |
| 105 | f"({within_tight_count}/{total}) " |
| 106 | f"fail={_format_observed_rate(fail_count, total)} " |
| 107 | f"allowed_fail={_format_allowed_rate(allowed_fail_rate, total)} " |
| 108 | f"abs>{max_atol}={_format_observed_rate(above_max_count, total)} " |
| 109 | f"allowed_above_max={_format_allowed_rate(max_fail_rate, total)} " |
| 110 | f"max_diff={max_diff:.6g} mean_diff={mean_diff:.6g} " |
| 111 | f"p99={p99:.6g} p999={p999:.6g}" |
| 112 | ) |
| 113 | print(msg) |
| 114 | if within_tight < 1.0: |
| 115 | warnings.warn(msg, stacklevel=2) |
| 116 | assert within_tight >= pass_rate, msg |
| 117 | assert above_max <= max_fail_rate, msg |
| 118 | |
| 119 | |
| 120 | def _assert_rel_error_budget( |
| 121 | actual: torch.Tensor, |
| 122 | expected: torch.Tensor, |
| 123 | *, |
| 124 | label: str, |
| 125 | tight_rtol: float, |
| 126 | max_rtol: float, |
| 127 | pass_rate: float, |
| 128 | max_fail_rate: float, |
| 129 | ) -> None: |
| 130 | rel = ( |
| 131 | (actual.float() - expected.float()).abs() |
| 132 | / expected.float().abs().clamp_min(1e-5) |
| 133 | ).flatten() |
| 134 | total = rel.numel() |
| 135 | within_tight_count = int((rel <= tight_rtol).sum().item()) |
| 136 | fail_count = total - within_tight_count |
| 137 | above_max_count = int((rel > max_rtol).sum().item()) |
| 138 | within_tight = within_tight_count / total |
| 139 | allowed_fail_rate = 1.0 - pass_rate |
| 140 | above_max = above_max_count / total |
| 141 | max_rel = rel.max().item() |
| 142 | mean_rel = rel.mean().item() |
| 143 | p99 = _quantile(rel, 0.99) |
| 144 | p999 = _quantile(rel, 0.999) |
| 145 | msg = ( |
| 146 | f"[rocm_fp8] {label}: rel<={tight_rtol} pass={within_tight:.4%} " |
| 147 | f"({within_tight_count}/{total}) " |
| 148 | f"fail={_format_observed_rate(fail_count, total)} " |
| 149 | f"allowed_fail={_format_allowed_rate(allowed_fail_rate, total)} " |
| 150 | f"rel>{max_rtol}={_format_observed_rate(above_max_count, total)} " |
| 151 | f"allowed_above_max={_format_allowed_rate(max_fail_rate, total)} " |
| 152 | f"max_rel={max_rel:.6g} mean_rel={mean_rel:.6g} " |
| 153 | f"p99={p99:.6g} p999={p999:.6g}" |
| 154 | ) |
| 155 | print(msg) |
| 156 | if within_tight < 1.0: |
| 157 | warnings.warn(msg, stacklevel=2) |
| 158 | assert within_tight >= pass_rate, msg |
| 159 | assert above_max <= max_fail_rate, msg |
| 160 | |
| 161 | |
| 162 | def _ref_paged_attention( |
| 163 | query: torch.Tensor, |
| 164 | key_cache: torch.Tensor, |
| 165 | value_cache: torch.Tensor, |
| 166 | *, |
| 167 | query_lens: list[int], |
| 168 | kv_lens: list[int], |
| 169 | block_tables: torch.Tensor, |
| 170 | scale: float, |
| 171 | ) -> torch.Tensor: |
| 172 | block_tables_np = block_tables.cpu().numpy() |
| 173 | _, block_size, num_kv_heads, head_size = key_cache.shape |
| 174 | |
| 175 | outputs = [] |
| 176 | start_idx = 0 |
| 177 | for i, query_len in enumerate(query_lens): |
| 178 | kv_len = kv_lens[i] |
| 179 | q = query[start_idx : start_idx + query_len] * scale |
| 180 | |
| 181 | num_kv_blocks = (kv_len + block_size - 1) // block_size |
| 182 | block_indices = block_tables_np[i, :num_kv_blocks] |
| 183 | |
| 184 | k = key_cache[block_indices].view(-1, num_kv_heads, head_size)[:kv_len] |
| 185 | v = value_cache[block_indices].view(-1, num_kv_heads, head_size)[:kv_len] |
| 186 | |
| 187 | if q.shape[1] != k.shape[1]: |
| 188 | repeat = q.shape[1] // k.shape[1] |
| 189 | k = torch.repeat_interleave(k, repeat, dim=1) |
| 190 | v = torch.repeat_interleave(v, repeat, dim=1) |
| 191 | |
| 192 | attn = torch.einsum("qhd,khd->hqk", q, k).float() |
| 193 | mask = torch.triu( |
| 194 | torch.ones(query_len, kv_len, device=q.device), |
| 195 | diagonal=kv_len - query_len + 1, |
| 196 | ).bool() |
| 197 | attn.masked_fill_(mask, float("-inf")) |
| 198 | attn = torch.softmax(attn, dim=-1).to(v.dtype) |
| 199 | outputs.append(torch.einsum("hqk,khd->qhd", attn, v)) |
| 200 | start_idx += query_len |
| 201 | |
| 202 | return torch.cat(outputs, dim=0) |
| 203 | |
| 204 | |
| 205 | def test_rocm_fp8_env_defaults(monkeypatch): |
| 206 | with monkeypatch.context() as mp: |
| 207 | for env_name in ENV_DEFAULTS: |
| 208 | mp.delenv(env_name, raising=False) |
| 209 | envs = _reload_envs() |
| 210 | for env_name, expected in ENV_DEFAULTS.items(): |
| 211 | assert getattr(envs, env_name) is expected |
| 212 | |
| 213 | |
| 214 | @pytest.mark.parametrize( |
| 215 | ("env_name", "raw_value", "expected"), |
| 216 | [ |
| 217 | ("VLLM_ROCM_FP8_PADDING", "0", False), |
| 218 | ("VLLM_ROCM_FP8_PADDING", "1", True), |
| 219 | ("VLLM_ROCM_USE_AITER_FP8BMM", "0", False), |
| 220 | ("VLLM_ROCM_USE_AITER_FP8BMM", "1", True), |
| 221 | ("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "0", False), |
| 222 | ("VLLM_ROCM_FP8_MFMA_PAGE_ATTN", "1", True), |
| 223 | ], |
| 224 | ) |
| 225 | def test_rocm_fp8_env_overrides(env_name, raw_value, expected, monkeypatch): |
| 226 | with monkeypatch.context() as mp: |
| 227 | mp.setenv(env_name, raw_value) |
| 228 | envs = _reload_envs() |
| 229 | assert getattr(envs, env_name) is expected |
| 230 | |
| 231 | |
| 232 | @fp8_only |
| 233 | def test_maybe_pad_fp8_weight_respects_env(monkeypatch): |
| 234 | from vllm.model_executor.layers.quantization.utils.fp8_utils import ( |
| 235 | _maybe_pad_fp8_weight, |
| 236 | ) |
| 237 | |
| 238 | weight = torch.randn(64, 512, device=DEVICE).to(FP8_DTYPE) |
| 239 | |
| 240 | with monkeypatch.context() as mp: |
| 241 | mp.setenv("VLLM_ROCM_FP8_PADDING", "1") |
| 242 | _reload_envs() |
| 243 | padded = _maybe_pad_fp8_weight(weight) |
| 244 | assert padded.shape == weight.shape |
| 245 | assert padded.dtype == weight.dtype |
| 246 | assert padded.data_ptr() != weight.data_ptr() |
| 247 | torch.testing.assert_close(padded.float(), weight.float()) |
| 248 | |
| 249 | with monkeypatch.context() as mp: |
| 250 | mp.setenv("VLLM_ROCM_FP8_PADDING", "0") |
| 251 | _reload_envs() |
| 252 | unpadded = _maybe_pad_fp8_weight(weight) |
| 253 | assert unpadded.data_ptr() == weight.data_ptr() |
| 254 | |
| 255 | |
| 256 | @fp8_only |
| 257 | @pytest.mark.parametrize( |
| 258 | ("use_aiter", "fp8bmm_enabled", "expected"), |
| 259 | [ |
| 260 | (True, True, True), |
| 261 | (True, False, False), |
| 262 | (False, True, False), |
| 263 | ], |
| 264 | ) |
| 265 | def test_aiter_fp8bmm_enabled_api_respects_env( |
| 266 | use_aiter, fp8bmm_enabled, expected, monkeypatch |
| 267 | ): |
| 268 | from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops |
| 269 | |
| 270 | assert is_aiter_found_and_supported() |
| 271 | |
| 272 | with monkeypatch.context() as mp: |
| 273 | mp.setenv("VLLM_ROCM_USE_AITER", "1" if use_aiter else "0") |
| 274 | mp.setenv("VLLM_ROCM_USE_AITER_FP8BMM", "1" if fp8bmm_enabled else "0") |
| 275 | _reload_envs() |
| 276 | rocm_aiter_ops.refresh_env_variables() |
| 277 | assert rocm_aiter_ops.is_fp8bmm_enabled() is expected |
| 278 | |
| 279 | |
| 280 | @fp8_only |
| 281 | @pytest.mark.parametrize( |
| 282 | ("shape", "tight_atol", "max_atol", "pass_rate"), SCALED_MM_CASES |
| 283 | ) |
| 284 | def test_fp8_scaled_mm_matches_dequantized_reference( |
| 285 | shape, tight_atol, max_atol, pass_rate |
| 286 | ): |
| 287 | torch.manual_seed(0) |
| 288 | |
| 289 | batch, m_dim, k_dim = shape |
| 290 | n_dim = k_dim |
| 291 | a = torch.randn(batch * m_dim, k_dim, device=DEVICE) |
| 292 | b = torch.randn(n_dim, k_dim, device=DEVICE) |
| 293 | |
| 294 | a_fp8, scale_a = ref_dynamic_per_tensor_fp8_quant(a) |
| 295 | b_fp8, scale_b = ref_dynamic_per_tensor_fp8_quant(b) |
| 296 | out = torch._scaled_mm( |
| 297 | a_fp8, |
| 298 | b_fp8.t(), |
| 299 | out_dtype=torch.bfloat16, |
| 300 | scale_a=scale_a, |
| 301 | scale_b=scale_b, |
| 302 | ) |
| 303 | |
| 304 | a_dq = a_fp8.float() * scale_a.float() |
| 305 | b_dq = b_fp8.float() * scale_b.float() |
| 306 | ref = torch.mm(a_dq, b_dq.t()) |
| 307 | |
| 308 | assert out.shape == (batch * m_dim, n_dim) |
| 309 | assert out.dtype == torch.bfloat16 |
| 310 | assert not torch.isnan(out).any() |
| 311 | |
| 312 | _assert_abs_error_budget( |
| 313 | out.float(), |
| 314 | ref, |
| 315 | label=f"scaled_mm shape={shape}", |
| 316 | tight_atol=tight_atol, |
| 317 | max_atol=max_atol, |
| 318 | pass_rate=pass_rate, |
| 319 | ) |
| 320 | |
| 321 | |
| 322 | @fp8_only |
| 323 | def test_fp8_scaled_mm_determinism(): |
| 324 | torch.manual_seed(5) |
| 325 | |
| 326 | m_dim, k_dim, n_dim = 64, 128, 128 |
| 327 | a = torch.randn(m_dim, k_dim, device=DEVICE) |
| 328 | b = torch.randn(n_dim, k_dim, device=DEVICE) |
| 329 | a_fp8, scale_a = ref_dynamic_per_tensor_fp8_quant(a) |
| 330 | b_fp8, scale_b = ref_dynamic_per_tensor_fp8_quant(b) |
| 331 | |
| 332 | def run_gemm(): |
| 333 | return torch._scaled_mm( |
| 334 | a_fp8, |
| 335 | b_fp8.t(), |
| 336 | out_dtype=torch.bfloat16, |
| 337 | scale_a=scale_a, |
| 338 | scale_b=scale_b, |
| 339 | ) |
| 340 | |
| 341 | _assert_deterministic(run_gemm, n_runs=4) |
| 342 | |
| 343 | |
| 344 | @fp8_only |
| 345 | def test_aiter_fp8_kv_attention_matches_bf16_reference(): |
| 346 | import aiter |
| 347 | |
| 348 | from vllm._aiter_ops import is_aiter_found_and_supported |
| 349 | from vllm.v1.attention.backends.rocm_aiter_fa import cp_mha_gather_cache |
| 350 | |
| 351 | assert is_aiter_found_and_supported() |
| 352 | |
| 353 | set_random_seed(10) |
| 354 | |
| 355 | num_q_heads, num_kv_heads = 8, 8 |
| 356 | head_size = 128 |
| 357 | query_len, kv_len = 4, 128 |
| 358 | block_size = 16 |
| 359 | num_blocks = 2048 |
| 360 | scale = head_size**-0.5 |
| 361 | |
| 362 | query = torch.randn( |
| 363 | query_len, num_q_heads, head_size, dtype=torch.bfloat16, device=DEVICE |
| 364 | ) |
| 365 | key_cache_fp8 = torch.clamp( |
| 366 | torch.randn(num_blocks, block_size, num_kv_heads, head_size, device=DEVICE), |
| 367 | min=-1.0, |
| 368 | max=1.0, |
| 369 | ).to(FP8_DTYPE) |
| 370 | value_cache_fp8 = torch.clamp( |
| 371 | torch.randn(num_blocks, block_size, num_kv_heads, head_size, device=DEVICE), |
| 372 | min=-1.0, |
| 373 | max=1.0, |
| 374 | ).to(FP8_DTYPE) |
| 375 | |
| 376 | cu_query_lens = torch.tensor([0, query_len], dtype=torch.int32, device=DEVICE) |
| 377 | cu_seq_lens = torch.tensor([0, kv_len], dtype=torch.int32, device=DEVICE) |
| 378 | block_tables = torch.randint( |
| 379 | 0, |
| 380 | num_blocks, |
| 381 | (1, (kv_len + block_size - 1) // block_size), |
| 382 | dtype=torch.int32, |
| 383 | device=DEVICE, |
| 384 | ) |
| 385 | token_to_batch = torch.zeros(kv_len, dtype=torch.int32, device=DEVICE) |
| 386 | seq_starts = torch.zeros(1, dtype=torch.int32, device=DEVICE) |
| 387 | gathered_key = torch.empty( |
| 388 | kv_len, num_kv_heads, head_size, dtype=torch.bfloat16, device=DEVICE |
| 389 | ) |
| 390 | gathered_value = torch.empty_like(gathered_key) |
| 391 | |
| 392 | cp_mha_gather_cache( |
| 393 | key_cache=key_cache_fp8, |
| 394 | value_cache=value_cache_fp8, |
| 395 | key=gathered_key, |
| 396 | value=gathered_value, |
| 397 | block_tables=block_tables, |
| 398 | k_scales=torch.ones(1, dtype=torch.float32, device=DEVICE), |
| 399 | v_scales=torch.ones(1, dtype=torch.float32, device=DEVICE), |
| 400 | cu_seqlens_kv=cu_seq_lens, |
| 401 | token_to_batch=token_to_batch, |
| 402 | seq_starts=seq_starts, |
| 403 | dequant=True, |
| 404 | kv_cache_layout="NHD", |
| 405 | total_tokens=kv_len, |
| 406 | ) |
| 407 | |
| 408 | output = torch.empty_like(query) |
| 409 | aiter.flash_attn_varlen_func( |
| 410 | q=query, |
| 411 | k=gathered_key, |
| 412 | v=gathered_value, |
| 413 | cu_seqlens_q=cu_query_lens, |
| 414 | cu_seqlens_k=cu_seq_lens, |
| 415 | max_seqlen_q=query_len, |
| 416 | max_seqlen_k=kv_len, |
| 417 | min_seqlen_q=1, |
| 418 | dropout_p=0.0, |
| 419 | softmax_scale=scale, |
| 420 | causal=True, |
| 421 | window_size=(-1, -1), |
| 422 | alibi_slopes=None, |
| 423 | return_lse=False, |
| 424 | out=output, |
| 425 | ) |
| 426 | |
| 427 | ref = _ref_paged_attention( |
| 428 | query=query, |
| 429 | key_cache=key_cache_fp8.to(torch.bfloat16), |
| 430 | value_cache=value_cache_fp8.to(torch.bfloat16), |
| 431 | query_lens=[query_len], |
| 432 | kv_lens=[kv_len], |
| 433 | block_tables=block_tables, |
| 434 | scale=scale, |
| 435 | ) |
| 436 | |
| 437 | _assert_abs_error_budget( |
| 438 | output.float(), |
| 439 | ref.float(), |
| 440 | label="aiter_fp8_kv_attention", |
| 441 | tight_atol=0.15, |
| 442 | max_atol=0.25, |
| 443 | pass_rate=0.98, |
| 444 | ) |
| 445 | |
| 446 | |
| 447 | @fp8_only |
| 448 | @pytest.mark.parametrize("shape", [(128, 256), (512, 1024), (1024, 4096)]) |
| 449 | def test_fp8_per_tensor_quant_matches_reference(shape): |
| 450 | torch.manual_seed(0) |
| 451 | |
| 452 | x = torch.randn(*shape, dtype=torch.bfloat16, device=DEVICE) |
| 453 | |
| 454 | out, scale = ops.scaled_fp8_quant(x) |
| 455 | ref_out, ref_scale = ref_dynamic_per_tensor_fp8_quant(x) |
| 456 | |
| 457 | assert out.shape == x.shape |
| 458 | assert out.dtype == FP8_DTYPE |
| 459 | |
| 460 | torch.testing.assert_close(scale, ref_scale, rtol=1e-3, atol=1e-4) |
| 461 | code_match_rate = ( |
| 462 | (out.view(torch.uint8) == ref_out.view(torch.uint8)).float().mean().item() |
| 463 | ) |
| 464 | assert code_match_rate >= 0.999, ( |
| 465 | f"FP8 codes match reference only {code_match_rate:.4%} (< 99.9%)" |
| 466 | ) |
| 467 | |
| 468 | dequant = out.float() * scale.float() |
| 469 | _assert_rel_error_budget( |
| 470 | dequant, |
| 471 | x.float(), |
| 472 | label=f"per_tensor_quant shape={shape}", |
| 473 | tight_rtol=0.06, |
| 474 | max_rtol=0.60, |
| 475 | pass_rate=0.99, |
| 476 | max_fail_rate=1e-4, |
| 477 | ) |
| 478 | |
| 479 | |
| 480 | @fp8_only |
| 481 | def test_aiter_group_fp8_quant_roundtrip(): |
| 482 | from aiter import dtypes |
| 483 | |
| 484 | from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops |
| 485 | |
| 486 | assert is_aiter_found_and_supported() |
| 487 | |
| 488 | torch.manual_seed(0) |
| 489 | |
| 490 | m_dim, n_dim = 64, 4096 |
| 491 | group_size = 128 |
| 492 | x = torch.randn(m_dim, n_dim, dtype=torch.bfloat16, device=DEVICE) |
| 493 | |
| 494 | out, scales = rocm_aiter_ops.group_fp8_quant(x, group_size) |
| 495 | assert out.dtype == dtypes.fp8 |
| 496 | assert scales.dtype == torch.float32 |
| 497 | |
| 498 | scales_expanded = scales.repeat_interleave(group_size, dim=1)[:, :n_dim] |
| 499 | dequant = out.float() * scales_expanded |
| 500 | _assert_rel_error_budget( |
| 501 | dequant, |
| 502 | x.float(), |
| 503 | label="aiter_group_fp8_quant", |
| 504 | tight_rtol=0.06, |
| 505 | max_rtol=0.30, |
| 506 | pass_rate=0.99, |
| 507 | max_fail_rate=1e-4, |
| 508 | ) |
| 509 | |
| 510 | |
| 511 | @fp8_only |
| 512 | def test_fp8_per_token_quant_matches_reference(): |
| 513 | torch.manual_seed(0) |
| 514 | |
| 515 | x = torch.randn(32, 4096, dtype=torch.bfloat16, device=DEVICE) |
| 516 | |
| 517 | out, scale = ops.scaled_fp8_quant(x, scale=None, use_per_token_if_dynamic=True) |
| 518 | ref_out, ref_scale = ref_dynamic_per_token_quant(x, FP8_DTYPE) |
| 519 | |
| 520 | assert out.shape == x.shape |
| 521 | assert out.dtype == FP8_DTYPE |
| 522 | assert scale.shape == (x.shape[0], 1) |
| 523 | |
| 524 | torch.testing.assert_close(scale, ref_scale, rtol=1e-3, atol=1e-4) |
| 525 | code_match_rate = ( |
| 526 | (out.view(torch.uint8) == ref_out.view(torch.uint8)).float().mean().item() |
| 527 | ) |
| 528 | assert code_match_rate >= 0.999, ( |
| 529 | f"FP8 codes match reference only {code_match_rate:.4%} (< 99.9%)" |
| 530 | ) |
| 531 | |
| 532 | dequant = out.float() * scale.float() |
| 533 | _assert_rel_error_budget( |
| 534 | dequant, |
| 535 | x.float(), |
| 536 | label="per_token_quant", |
| 537 | tight_rtol=0.06, |
| 538 | max_rtol=0.30, |
| 539 | pass_rate=0.99, |
| 540 | max_fail_rate=1e-4, |
| 541 | ) |