Merge branch 'main' into habana_upstream

benchmarks/kernels/benchmark_layernorm.py

@@ -3,8 +3,8 @@
 import torch
 
 from vllm.model_executor.layers.layernorm import RMSNorm
-from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        seed_everything)
+from vllm.platforms import current_platform
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
 
 
 @torch.inference_mode()
@@ -16,7 +16,7 @@ def main(num_tokens: int,
          do_profile: bool = False,
          num_warmup_iters: int = 5,
          num_iters: int = 100) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device("cuda")
 
     layer = RMSNorm(hidden_size).to(dtype=dtype)

benchmarks/kernels/benchmark_moe.py

@@ -10,7 +10,8 @@
 from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
-from vllm.utils import FlexibleArgumentParser, seed_everything
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
 
 
 class BenchmarkConfig(TypedDict):
@@ -167,7 +168,7 @@ class BenchmarkWorker:
 
     def __init__(self, seed: int) -> None:
         torch.set_default_device("cuda")
-        seed_everything(seed)
+        current_platform.seed_everything(seed)
         self.seed = seed
 
     def benchmark(
@@ -181,7 +182,7 @@ def benchmark(
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
     ) -> Tuple[Dict[str, int], float]:
-        seed_everything(self.seed)
+        current_platform.seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
                                          use_int8_w8a16=use_int8_w8a16,
                                          use_fp8_w8a8=use_fp8_w8a8)

benchmarks/kernels/benchmark_paged_attention.py

@@ -5,8 +5,9 @@
 import torch
 
 from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        create_kv_caches_with_random, seed_everything)
+                        create_kv_caches_with_random)
 
 NUM_BLOCKS = 1024
 PARTITION_SIZE = 512
@@ -28,7 +29,7 @@ def main(
     device: str = "cuda",
     kv_cache_dtype: Optional[str] = None,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
 
     scale = float(1.0 / (head_size**0.5))
     query = torch.empty(num_seqs,

benchmarks/kernels/benchmark_quant.py

@@ -3,8 +3,8 @@
 import torch
 
 from vllm import _custom_ops as ops
-from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        seed_everything)
+from vllm.platforms import current_platform
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
 
 
 @torch.inference_mode()
@@ -17,7 +17,7 @@ def main(num_tokens: int,
          do_profile: bool = False,
          num_warmup_iters: int = 5,
          num_iters: int = 100) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device("cuda")
 
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)

benchmarks/kernels/benchmark_rope.py

@@ -6,7 +6,8 @@
 
 from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
                                                          get_rope)
-from vllm.utils import FlexibleArgumentParser, seed_everything
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
 
 
 def benchmark_rope_kernels_multi_lora(
@@ -22,7 +23,7 @@ def benchmark_rope_kernels_multi_lora(
     max_position: int = 8192,
     base: int = 10000,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     if rotary_dim is None:
         rotary_dim = head_size

tests/kernels/test_activation.py

@@ -8,7 +8,7 @@
 from vllm.model_executor.layers.activation import (FastGELU, FatreluAndMul,
                                                    GeluAndMul, NewGELU,
                                                    QuickGELU, SiluAndMul)
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 
 from .allclose_default import get_default_atol, get_default_rtol
 
@@ -37,7 +37,7 @@ def test_act_and_mul(
     seed: int,
     device: str,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     x = torch.randn(num_tokens, 2 * d, dtype=dtype)
     if activation == "silu":
@@ -85,7 +85,7 @@ def test_activation(
     seed: int,
     device: str,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     x = torch.randn(num_tokens, d, dtype=dtype)
     layer = activation[0]()

tests/kernels/test_attention.py

@@ -7,7 +7,7 @@
 from tests.kernels.utils import opcheck
 from vllm import _custom_ops as ops
 from vllm.platforms import current_platform
-from vllm.utils import get_max_shared_memory_bytes, seed_everything
+from vllm.utils import get_max_shared_memory_bytes
 
 from .allclose_default import get_default_atol, get_default_rtol
 
@@ -144,7 +144,7 @@ def test_paged_attention(
             or (version == "rocm" and head_size not in (64, 128))):
         pytest.skip()
 
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     scale = float(1.0 / (head_size**0.5))
     num_query_heads, num_kv_heads = num_heads
@@ -382,7 +382,7 @@ def test_multi_query_kv_attention(
     seed: int,
     device: str,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
     # As the xformers library is already tested with its own tests, we can use

tests/kernels/test_awq_triton.py

@@ -7,7 +7,7 @@
 
 from vllm.model_executor.layers.quantization.awq_triton import (
     AWQ_TRITON_SUPPORTED_GROUP_SIZES, awq_dequantize_triton, awq_gemm_triton)
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 
 device = "cuda"
 
@@ -80,7 +80,7 @@ def test_dequantize(qweight_rows, qweight_cols, group_size):
     zeros_cols = qweight_cols
     zeros_dtype = torch.int32
 
-    seed_everything(0)
+    current_platform.seed_everything(0)
 
     qweight = torch.randint(0,
                             torch.iinfo(torch.int32).max,
@@ -134,7 +134,7 @@ def test_gemm(N, K, M, splitK, group_size):
     qzeros_rows = scales_rows
     qzeros_cols = qweight_cols
 
-    seed_everything(0)
+    current_platform.seed_everything(0)
 
     input = torch.rand((input_rows, input_cols),
                        dtype=input_dtype,

tests/kernels/test_blocksparse_attention.py

@@ -8,7 +8,7 @@
 from vllm.attention.ops.blocksparse_attention.interface import (
     LocalStridedBlockSparseAttn)
 from vllm.platforms import current_platform
-from vllm.utils import get_max_shared_memory_bytes, seed_everything
+from vllm.utils import get_max_shared_memory_bytes
 
 from .allclose_default import get_default_atol, get_default_rtol
 
@@ -173,7 +173,7 @@ def test_paged_attention(
     blocksparse_block_size: int,
     blocksparse_head_sliding_step: int,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     scale = float(1.0 / (head_size**0.5))
     num_query_heads, num_kv_heads = num_heads
@@ -384,7 +384,7 @@ def test_varlen_blocksparse_attention_prefill(
     seed: int,
     device: str,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     # MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
     # As the xformers library is already tested with its own tests, we can use

tests/kernels/test_cache.py

@@ -6,7 +6,7 @@
 
 from tests.kernels.utils import DEFAULT_OPCHECK_TEST_UTILS, opcheck
 from vllm import _custom_ops as ops
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 
 COPYING_DIRECTION = [('cuda', 'cpu'), ('cuda', 'cuda'), ('cpu', 'cuda')]
 DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -56,7 +56,7 @@ def test_copy_blocks(
 ) -> None:
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     # Generate random block mappings where each source block is mapped to two
     # destination blocks.
@@ -132,7 +132,7 @@ def test_reshape_and_cache(
 ) -> None:
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     # Create a random slot mapping.
     num_slots = block_size * num_blocks
@@ -224,7 +224,7 @@ def test_reshape_and_cache_flash(
     device: str,
     kv_cache_dtype: str,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
 
     # Create a random slot mapping.
@@ -339,7 +339,7 @@ def test_swap_blocks(
     if kv_cache_dtype == "fp8" and head_size % 16:
         pytest.skip()
 
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
 
     src_device = device if direction[0] == "cuda" else 'cpu'
     dst_device = device if direction[1] == "cuda" else 'cpu'
@@ -408,7 +408,7 @@ def test_fp8_e4m3_conversion(
     seed: int,
     device: str,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
 
     low = -224.0
     high = 224.0

tests/kernels/test_causal_conv1d.py

@@ -9,7 +9,7 @@
 from vllm.attention.backends.utils import PAD_SLOT_ID
 from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
     causal_conv1d_fn, causal_conv1d_update)
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 
 
 def causal_conv1d_ref(
@@ -70,7 +70,7 @@ def causal_conv1d_update_ref(x,
     bias: (dim,)
     cache_seqlens: (batch,), dtype int32.
         If not None, the conv_state is treated as a circular buffer.
-        The conv_state will be updated by copying x to the 
+        The conv_state will be updated by copying x to the
         conv_state starting at the index
         @cache_seqlens % state_len before performing the convolution.
 
@@ -161,7 +161,7 @@ def test_causal_conv1d(batch, dim, seqlen, width, has_bias, silu_activation,
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    seed_everything(0)
+    current_platform.seed_everything(0)
     x = torch.randn(batch, dim, seqlen, device=device,
                     dtype=itype).contiguous()
 
@@ -223,7 +223,7 @@ def test_causal_conv1d_update(dim, width, seqlen, has_bias, silu_activation,
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    seed_everything(0)
+    current_platform.seed_everything(0)
     batch = 2
     x = torch.randn(batch, dim, seqlen, device=device, dtype=itype)
     x_ref = x.clone()
@@ -270,7 +270,7 @@ def test_causal_conv1d_update_with_batch_gather(with_padding, dim, width,
         rtol, atol = 1e-2, 5e-2
 
     # set seed
-    seed_everything(0)
+    current_platform.seed_everything(0)
 
     batch_size = 3
     padding = 5 if with_padding else 0
@@ -343,7 +343,7 @@ def test_causal_conv1d_varlen(with_padding, dim, seqlen, width, has_bias,
     if itype == torch.bfloat16:
         rtol, atol = 1e-2, 5e-2
     # set seed
-    seed_everything(0)
+    current_platform.seed_everything(0)
     seqlens = []
     batch_size = 4
     if seqlen < 10:

tests/kernels/test_flash_attn.py

@@ -3,7 +3,7 @@
 import pytest
 import torch
 
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 from vllm.vllm_flash_attn import (flash_attn_varlen_func,
                                   flash_attn_with_kvcache)
 
@@ -91,7 +91,7 @@ def test_flash_attn_with_paged_kv(
     sliding_window: Optional[int],
 ) -> None:
     torch.set_default_device("cuda")
-    seed_everything(0)
+    current_platform.seed_everything(0)
     num_seqs = len(kv_lens)
     num_query_heads = num_heads[0]
     num_kv_heads = num_heads[1]
@@ -161,7 +161,7 @@ def test_varlen_with_paged_kv(
     num_blocks: int,
 ) -> None:
     torch.set_default_device("cuda")
-    seed_everything(0)
+    current_platform.seed_everything(0)
     num_seqs = len(seq_lens)
     query_lens = [x[0] for x in seq_lens]
     kv_lens = [x[1] for x in seq_lens]

tests/kernels/test_flashinfer.py

@@ -4,7 +4,7 @@
 import pytest
 import torch
 
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 
 NUM_HEADS = [(16, 16), (32, 8), (64, 8), (6, 1)]
 HEAD_SIZES = [128, 256]
@@ -84,7 +84,7 @@ def test_flashinfer_decode_with_paged_kv(
     soft_cap: Optional[float],
 ) -> None:
     torch.set_default_device("cuda")
-    seed_everything(0)
+    current_platform.seed_everything(0)
     num_seqs = len(kv_lens)
     num_query_heads = num_heads[0]
     num_kv_heads = num_heads[1]
@@ -170,7 +170,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
                                           block_size: int,
                                           soft_cap: Optional[float]) -> None:
     torch.set_default_device("cuda")
-    seed_everything(0)
+    current_platform.seed_everything(0)
     num_seqs = len(seq_lens)
     query_lens = [x[0] for x in seq_lens]
     kv_lens = [x[1] for x in seq_lens]
@@ -268,7 +268,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
         head_size: int, dtype: torch.dtype, block_size: int,
         soft_cap: Optional[float]) -> None:
     torch.set_default_device("cuda")
-    seed_everything(0)
+    current_platform.seed_everything(0)
     num_seqs = len(seq_lens)
     query_lens = [x[0] for x in seq_lens]
     kv_lens = [x[1] for x in seq_lens]
@@ -381,7 +381,7 @@ def test_flashinfer_decode_with_paged_fp8_kv(
 ) -> None:
     # test doesn't work for num_heads = (16,16)
     torch.set_default_device("cuda")
-    seed_everything(0)
+    current_platform.seed_everything(0)
     num_seqs = len(kv_lens)
     num_query_heads = num_heads[0]
     num_kv_heads = num_heads[1]