| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """XPU-only batch-invariance tests for norm, collectives, and sampler behavior.""" |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | from vllm import envs |
| 9 | from vllm.config import CacheConfig, LoadConfig, ModelConfig, VllmConfig |
| 10 | from vllm.distributed.device_communicators.xpu_communicator import XpuCommunicator |
| 11 | from vllm.platforms import current_platform |
| 12 | |
| 13 | pytestmark = pytest.mark.skipif( |
| 14 | not current_platform.is_xpu(), reason="XPU batch-invariance tests" |
| 15 | ) |
| 16 | |
| 17 | |
| 18 | @pytest.fixture |
| 19 | def tiny_qwen3_moe_path(tmp_path): |
| 20 | """Create a local Qwen3-MoE config for dummy-loading tests.""" |
| 21 | from transformers import Qwen3MoeConfig |
| 22 | |
| 23 | config = Qwen3MoeConfig( |
| 24 | architectures=["Qwen3MoeForCausalLM"], |
| 25 | vocab_size=256, |
| 26 | hidden_size=512, |
| 27 | intermediate_size=1024, |
| 28 | moe_intermediate_size=512, |
| 29 | num_hidden_layers=2, |
| 30 | num_attention_heads=4, |
| 31 | num_key_value_heads=2, |
| 32 | head_dim=128, |
| 33 | num_experts=4, |
| 34 | num_experts_per_tok=2, |
| 35 | max_position_embeddings=512, |
| 36 | bos_token_id=1, |
| 37 | eos_token_id=2, |
| 38 | tie_word_embeddings=False, |
| 39 | ) |
| 40 | config.save_pretrained(tmp_path) |
| 41 | return tmp_path |
| 42 | |
| 43 | |
| 44 | @pytest.mark.parametrize("source", ["argument", "checkpoint"]) |
| 45 | def test_rejects_weight_quantization(tiny_qwen3_moe_path, source): |
| 46 | """Reject both explicit and checkpoint-inferred quantization before loading.""" |
| 47 | from transformers import Qwen3MoeConfig |
| 48 | |
| 49 | if source == "checkpoint": |
| 50 | config = Qwen3MoeConfig.from_pretrained(tiny_qwen3_moe_path) |
| 51 | config.quantization_config = { |
| 52 | "quant_method": "fp8", |
| 53 | "activation_scheme": "dynamic", |
| 54 | } |
| 55 | config.save_pretrained(tiny_qwen3_moe_path) |
| 56 | |
| 57 | model_config = ModelConfig( |
| 58 | model=str(tiny_qwen3_moe_path), |
| 59 | skip_tokenizer_init=True, |
| 60 | dtype="bfloat16", |
| 61 | quantization="fp8" if source == "argument" else None, |
| 62 | ) |
| 63 | assert model_config.quantization == "fp8" |
| 64 | with pytest.raises(ValueError, match="supports only unquantized models"): |
| 65 | VllmConfig( |
| 66 | model_config=model_config, |
| 67 | load_config=LoadConfig(load_format="dummy"), |
| 68 | ) |
| 69 | |
| 70 | |
| 71 | @pytest.mark.parametrize("cache_dtype", ["fp8", "int8_per_token_head", "nvfp4"]) |
| 72 | def test_rejects_quantized_kv_cache(cache_dtype): |
| 73 | with pytest.raises(ValueError, match="does not support quantized KV caches"): |
| 74 | VllmConfig(cache_config=CacheConfig(cache_dtype=cache_dtype)) |
| 75 | |
| 76 | |
| 77 | @pytest.mark.parametrize("cache_dtype", ["auto", "float16", "bfloat16"]) |
| 78 | def test_accepts_unquantized_kv_cache(cache_dtype): |
| 79 | config = VllmConfig(cache_config=CacheConfig(cache_dtype=cache_dtype)) |
| 80 | assert config.cache_config.cache_dtype == cache_dtype |
| 81 | |
| 82 | |
| 83 | def test_quantized_kv_cache_allowed_without_batch_invariance(monkeypatch): |
| 84 | monkeypatch.setattr(envs, "VLLM_BATCH_INVARIANT", False) |
| 85 | config = VllmConfig(cache_config=CacheConfig(cache_dtype="fp8")) |
| 86 | assert config.cache_config.cache_dtype == "fp8" |
| 87 | |
| 88 | |
| 89 | # These dimensions and seeds locally reproduce non-invariant norm mismatches. |
| 90 | # They make the invariant equality assertion exercise a discriminating case. |
| 91 | @pytest.mark.parametrize( |
| 92 | ( |
| 93 | "norm_name", |
| 94 | "batch_size", |
| 95 | "hidden_size", |
| 96 | "position", |
| 97 | "weight_seed", |
| 98 | "input_seed", |
| 99 | "input_pattern", |
| 100 | ), |
| 101 | [ |
| 102 | ("rms", 1024, 16384, 512, 7, 42, "normal"), |
| 103 | ("gemma", 257, 8192, 255, 1, 2, "normal"), |
| 104 | ("rms", 5, 8192, 3, 7, 46, "scaled"), |
| 105 | ("gemma", 5, 8192, 3, 1, 46, "scaled"), |
| 106 | ], |
| 107 | ) |
| 108 | def test_residual_norm_preserves_batch_invariance( |
| 109 | norm_name, |
| 110 | batch_size, |
| 111 | hidden_size, |
| 112 | position, |
| 113 | weight_seed, |
| 114 | input_seed, |
| 115 | input_pattern, |
| 116 | monkeypatch, |
| 117 | default_vllm_config, |
| 118 | ): |
| 119 | """Residual norm output must not change when its row changes batch position.""" |
| 120 | from vllm.model_executor.determinism.batch_invariant import init_batch_invariance |
| 121 | from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm |
| 122 | |
| 123 | monkeypatch.setattr(envs, "VLLM_BATCH_INVARIANT", True) |
| 124 | init_batch_invariance() |
| 125 | norm_cls = RMSNorm if norm_name == "rms" else GemmaRMSNorm |
| 126 | norm = norm_cls(hidden_size, eps=1e-6).to(device="xpu", dtype=torch.bfloat16) |
| 127 | with torch.no_grad(): |
| 128 | torch.manual_seed(weight_seed) |
| 129 | norm.weight.uniform_(-0.5, 0.5) |
| 130 | |
| 131 | if input_pattern == "scaled": |
| 132 | # Heterogeneous columns expose native mean's batch-dependent rounding. |
| 133 | generator = torch.Generator(device="cpu").manual_seed(input_seed) |
| 134 | scale = torch.randn(hidden_size, generator=generator).exp() |
| 135 | x = (torch.randn(batch_size, hidden_size, generator=generator) * scale).to( |
| 136 | device="xpu", dtype=torch.bfloat16 |
| 137 | ) |
| 138 | residual = ( |
| 139 | torch.randn(batch_size, hidden_size, generator=generator) * scale |
| 140 | ).to(device="xpu", dtype=torch.bfloat16) |
| 141 | else: |
| 142 | torch.manual_seed(input_seed) |
| 143 | x = torch.randn(batch_size, hidden_size, device="xpu", dtype=torch.bfloat16) |
| 144 | residual = torch.randn_like(x) |
| 145 | |
| 146 | single_output, single_residual = norm( |
| 147 | x[position : position + 1].clone(), |
| 148 | residual[position : position + 1].clone(), |
| 149 | ) |
| 150 | batch_output, batch_residual = norm(x.clone(), residual.clone()) |
| 151 | |
| 152 | assert torch.equal(single_output, batch_output[position : position + 1]) |
| 153 | assert torch.equal(single_residual, batch_residual[position : position + 1]) |
| 154 | |
| 155 | |
| 156 | @pytest.mark.parametrize( |
| 157 | "operation", ["all_reduce", "reduce_scatter", "reduce_scatterv"] |
| 158 | ) |
| 159 | @pytest.mark.parametrize("rank", range(4)) |
| 160 | def test_collectives_preserve_rank_order_and_partition(monkeypatch, operation, rank): |
| 161 | """Cancellation exposes reordered sums; mock ranks require no extra XPUs.""" |
| 162 | from vllm.distributed.device_communicators import xpu_communicator |
| 163 | |
| 164 | inputs = [ |
| 165 | torch.full((8, 3), value, dtype=torch.float32, device="xpu") |
| 166 | for value in (1e20, -1e20, 3.0, 0.0) |
| 167 | ] |
| 168 | inputs[2].add_(torch.arange(24, device="xpu").reshape(8, 3)) |
| 169 | |
| 170 | def all_gather_single(output, input_, group): |
| 171 | torch.testing.assert_close(input_, inputs[rank].reshape(-1), rtol=0, atol=0) |
| 172 | output.copy_(torch.stack(inputs).reshape(-1)) |
| 173 | |
| 174 | monkeypatch.setattr(xpu_communicator.dist, "all_gather_single", all_gather_single) |
| 175 | communicator = XpuCommunicator.__new__(XpuCommunicator) |
| 176 | communicator.world_size = 4 |
| 177 | communicator.rank_in_group = rank |
| 178 | communicator.device_group = object() |
| 179 | expected = inputs[0].clone() |
| 180 | for value in inputs[1:]: |
| 181 | expected.add_(value) |
| 182 | assert torch.equal(expected, inputs[2]) |
| 183 | |
| 184 | if operation == "all_reduce": |
| 185 | actual = communicator.all_reduce(inputs[rank]) |
| 186 | elif operation == "reduce_scatter": |
| 187 | actual = communicator.reduce_scatter(inputs[rank], dim=0) |
| 188 | expected = expected.chunk(4)[rank] |
| 189 | else: |
| 190 | sizes = [1, 3, 2, 2] |
| 191 | actual = communicator.reduce_scatterv(inputs[rank], dim=0, sizes=sizes) |
| 192 | expected = expected.split(sizes)[rank] |
| 193 | torch.testing.assert_close(actual, expected, rtol=0, atol=0) |
| 194 | |
| 195 | |
| 196 | # This batch size, vocabulary size, and positions locally reproduce an |
| 197 | # unseeded non-invariant XPU sampler mismatch. |
| 198 | @pytest.mark.parametrize("position", [0, 128, 255]) |
| 199 | def test_seeded_sampler_preserves_full_distribution_and_rng(position, monkeypatch): |
| 200 | """Moving a request must preserve its distribution and per-request RNG.""" |
| 201 | from vllm.triton_utils import HAS_TRITON |
| 202 | from vllm.v1.sample.ops.topk_topp_sampler import TopKTopPSampler |
| 203 | |
| 204 | assert HAS_TRITON, "XPU sampling must exercise the Triton top-k/top-p path" |
| 205 | monkeypatch.setattr(envs, "VLLM_XPU_USE_SAMPLER_KERNEL", True) |
| 206 | sampler = TopKTopPSampler(logprobs_mode="processed_logprobs") |
| 207 | torch.manual_seed(42) |
| 208 | logits = torch.randn(256, 8192, device="xpu", dtype=torch.float32) |
| 209 | single_rng = torch.Generator(device="xpu").manual_seed(123) |
| 210 | batch_rng = torch.Generator(device="xpu").manual_seed(123) |
| 211 | k = torch.full((256,), 32, dtype=torch.int32, device="xpu") |
| 212 | p = torch.full((256,), 0.9, device="xpu") |
| 213 | for _ in range(3): |
| 214 | single_token, single_logprobs = sampler( |
| 215 | logits[position : position + 1].clone(), |
| 216 | {0: single_rng}, |
| 217 | k[position : position + 1], |
| 218 | p[position : position + 1], |
| 219 | ) |
| 220 | batch_tokens, batch_logprobs = sampler( |
| 221 | logits.clone(), {position: batch_rng}, k, p |
| 222 | ) |
| 223 | assert single_logprobs is not None and batch_logprobs is not None |
| 224 | torch.testing.assert_close( |
| 225 | single_logprobs[0], batch_logprobs[position], rtol=0, atol=0 |
| 226 | ) |
| 227 | assert single_token.item() == batch_tokens[position].item() |
| 228 | assert torch.equal(single_rng.get_state(), batch_rng.get_state()) |