move benchmark scripts to example

tests/test_infer/test_ops/cuda/test_kv_cache_memcpy.py

@@ -6,7 +6,7 @@
 from colossalai.kernel.kernel_loader import InferenceOpsLoader
 from colossalai.kernel.triton import copy_kv_to_blocked_cache
 from colossalai.utils import get_current_device
-from tests.test_infer.test_ops.triton.kernel_utils import generate_caches_and_block_tables_v2, mock_alloc_single_token
+from tests.test_infer.test_ops.triton.test_kvcache_copy import prepare_data
 
 try:
     import triton  # noqa
@@ -30,45 +30,6 @@
 HEAD_DIM = 4
 
 
-def prepare_data(
-    bsz,
-    num_kv_heads,
-    head_dim,
-    block_size,
-    max_num_blocks_per_seq,
-    same_context_len,
-    max_seq_len,
-    device,
-    dtype=torch.float32,
-):
-    # past_kv_seq_lengths in this test records the previous kv seq len
-    # (not incorporating the current input whose seq len is 1)
-    past_kv_seq_lengths = (
-        torch.tensor([max_seq_len - 1 for _ in range(bsz)], dtype=torch.int32, device=device)
-        if same_context_len
-        else torch.randint(low=1, high=max_seq_len - 1, size=(bsz,), dtype=torch.int32, device=device)
-    )
-    num_tokens = torch.sum(past_kv_seq_lengths).item()
-
-    kv_size = (num_tokens, 2 * num_kv_heads, head_dim)
-    kv_unpad = torch.empty(size=kv_size, dtype=dtype, device=device).normal_(mean=0.0, std=0.5)
-    k_unpad, v_unpad = torch.split(kv_unpad, [num_kv_heads, num_kv_heads], dim=-2)
-
-    k_cache, v_cache, block_tables = generate_caches_and_block_tables_v2(
-        k_unpad, v_unpad, past_kv_seq_lengths, bsz, max_num_blocks_per_seq, block_size, dtype=dtype, device=device
-    )
-    block_tables = block_tables.to(device=device)
-
-    new_k = torch.randn((bsz, 1, num_kv_heads, head_dim), dtype=dtype, device=device)
-    new_v = torch.randn((bsz, 1, num_kv_heads, head_dim), dtype=dtype, device=device)
-    # mock allocating blocks for the new k/v and update block tables
-    mock_alloc_single_token(block_tables, past_kv_seq_lengths, block_size)
-    # kv seq len = past kv seq len + seq len (1 during decoding stage)
-    kv_seq_lengths = past_kv_seq_lengths + 1
-
-    return new_k, new_v, k_cache, v_cache, kv_seq_lengths, block_tables
-
-
 @pytest.mark.skipif(not (HAS_TRITON and TRITON_CUDA_SUPPORT), reason="requires triton")
 @pytest.mark.parametrize("bsz", [4, 7, 32])
 @pytest.mark.parametrize("block_size", [16, 32, 64])
@@ -119,9 +80,6 @@ def test_copy_kv_to_caches(
     assert torch.equal(k_target, k_source)
     assert v_target.shape == v_source.shape
     assert torch.equal(v_target, v_source)
-    # target_torch = k_cache_copy[target_block_ids, :, offsets_in_block, :]
-    # assert target_torch.shape == source.shape
-    # assert torch.equal(target_torch, source)
 
 
 BATCH = 16

tests/test_infer/test_ops/triton/test_context_attn_unpad.py

@@ -1,9 +1,7 @@
 import pytest
 import torch
 from packaging import version
-from transformers.modeling_attn_mask_utils import AttentionMaskConverter
 
-from colossalai.inference.modeling.layers.attention import PagedAttention
 from colossalai.kernel.triton import context_attention_unpadded
 from colossalai.utils import get_current_device
 from tests.test_infer.test_ops.triton.kernel_utils import generate_caches_and_block_tables_v2, torch_attn_ref
@@ -116,102 +114,5 @@ def test_context_attention(
     assert torch.equal(v_cache_ref, v_cache_triton)
 
 
-BATCH = 16
-BLOCK_SIZE = 32
-SAME_LEN = True
-WARM_UPS = 10
-REPS = 100
-configs = [
-    triton.testing.Benchmark(
-        x_names=["KV_LEN"],
-        x_vals=[2**i for i in range(8, 13)],
-        # x_vals=[x for x in range(256, 8192, 256)],
-        line_arg="provider",
-        line_vals=["torch", "triton"],
-        line_names=["Torch", "Triton"],
-        styles=[("red", "-"), ("blue", "-")],
-        ylabel="ms",
-        plot_name=f"context_attn-block_size-{BLOCK_SIZE}-batch{BATCH}",
-        args={"bsz": BATCH, "block_size": BLOCK_SIZE, "same_context_len": SAME_LEN, "kv_group_num": 1},
-    )
-]
-
-
-@triton.testing.perf_report(configs)
-def bench_kernel(
-    bsz,
-    KV_LEN,
-    provider,
-    block_size: int,
-    kv_group_num: int,
-    same_context_len: bool,
-):
-    num_attn_heads = 16
-    max_num_blocks_per_seq = triton.cdiv(KV_LEN, block_size)
-    max_seq_len = block_size * max_num_blocks_per_seq
-
-    num_kv_heads = num_attn_heads // kv_group_num
-    assert isinstance(num_kv_heads, int) and num_kv_heads > 0, "Invalid number of kv heads."
-    dtype = torch.float16
-    device = get_current_device()
-
-    if same_context_len:
-        context_lengths = torch.tensor([max_seq_len for _ in range(bsz)], dtype=torch.int32, device=device)
-    else:
-        context_lengths = torch.randint(low=1, high=max_seq_len, size=(bsz,), dtype=torch.int32, device=device)
-    num_tokens = torch.sum(context_lengths).item()
-
-    qkv_size = (num_tokens, num_attn_heads + 2 * num_kv_heads, HEAD_DIM)
-    qkv_unpad = torch.empty(size=qkv_size, dtype=dtype, device=device).normal_(mean=0.0, std=0.5)
-    q_unpad, k_unpad, v_unpad = torch.split(qkv_unpad, [num_attn_heads, num_kv_heads, num_kv_heads], dim=-2)
-    q_unpad = q_unpad.contiguous()
-    k_cache_ref, v_cache_ref, block_tables = generate_caches_and_block_tables_v2(
-        k_unpad, v_unpad, context_lengths, bsz, max_num_blocks_per_seq, block_size, dtype, device
-    )
-    block_tables = block_tables.to(device=device)
-
-    quantiles = [0.5, 0.2, 0.8]
-    if provider == "torch":
-        q_padded = PagedAttention.pad_and_reshape(q_unpad, context_lengths, max_seq_len, num_attn_heads, HEAD_DIM)
-        k_padded = PagedAttention.pad_and_reshape(k_unpad, context_lengths, max_seq_len, num_kv_heads, HEAD_DIM)
-        v_padded = PagedAttention.pad_and_reshape(v_unpad, context_lengths, max_seq_len, num_kv_heads, HEAD_DIM)
-        q_padded, k_padded, v_padded = (
-            q_padded.to(device=device),
-            k_padded.to(device=device),
-            v_padded.to(device=device),
-        )
-        q_padded = q_padded.transpose(1, 2)
-        k_padded = PagedAttention.repeat_kv(k_padded.transpose(1, 2), kv_group_num)
-        v_padded = PagedAttention.repeat_kv(v_padded.transpose(1, 2), kv_group_num)
-        # This benchmark ignores the padding mask. *Only* use the-same-length inputs for benchmarkings
-        attn_mask = AttentionMaskConverter._make_causal_mask(
-            (bsz, max_seq_len), q_padded.dtype, q_padded.device, past_key_values_length=0
-        )
-        attn_mask = attn_mask.to(device=q_padded.device)
-        fn = lambda: torch_attn_ref(
-            q_padded,
-            k_padded,
-            v_padded,
-            attn_mask,
-            bsz,
-            max_seq_len,
-            max_seq_len,
-            num_attn_heads,
-            num_kv_heads,
-            HEAD_DIM,
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench(fn, warmup=WARM_UPS, rep=REPS, quantiles=quantiles)
-    if provider == "triton":
-        k_cache_triton = torch.zeros_like(k_cache_ref)
-        v_cache_triton = torch.zeros_like(v_cache_ref)
-        fn = lambda: context_attention_unpadded(
-            q_unpad, k_unpad, v_unpad, k_cache_triton, v_cache_triton, context_lengths, block_tables, block_size
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench(fn, warmup=WARM_UPS, rep=REPS, quantiles=quantiles)
-
-    return ms, min_ms, max_ms
-
-
 if __name__ == "__main__":
     test_context_attention(4, 32, 8, 16, 1, True)
-    # bench_kernel.run(save_path=".", print_data=True)

tests/test_infer/test_ops/triton/test_decoding_attn.py

@@ -128,94 +128,5 @@ def test_flash_decoding(
     assert torch.allclose(out_torch, out_triton, atol=1e-3, rtol=1e-4)
 
 
-BATCH = 16
-BLOCK_SIZE = 32
-SAME_LEN = True
-WARM_UPS = 10
-REPS = 100
-configs = [
-    triton.testing.Benchmark(
-        x_names=["KV_LEN"],
-        x_vals=[2**i for i in range(8, 14)],
-        # x_vals=[x for x in range(256, 8192, 256)],
-        line_arg="provider",
-        line_vals=["torch", "triton"],
-        line_names=["Torch", "Triton"],
-        styles=[("red", "-"), ("blue", "-")],
-        ylabel="ms",
-        plot_name=f"decoding-block_size-{BLOCK_SIZE}-batch{BATCH}",
-        args={"bsz": BATCH, "block_size": BLOCK_SIZE, "same_context_len": SAME_LEN, "kv_group_num": 1},
-    )
-]
-
-
-@triton.testing.perf_report(configs)
-def bench_kernel(
-    bsz,
-    KV_LEN,
-    provider,
-    block_size: int,
-    kv_group_num: int,
-    same_context_len: bool,
-):
-    num_attn_heads = 16
-    max_num_blocks_per_seq = triton.cdiv(KV_LEN, block_size)
-    max_seq_len = block_size * max_num_blocks_per_seq
-
-    num_kv_heads = num_attn_heads // kv_group_num
-    assert isinstance(num_kv_heads, int) and num_kv_heads > 0, "Invalid number of kv heads."
-    block_size * max_num_blocks_per_seq
-    dtype = torch.float16
-    device = get_current_device()
-
-    q, k_unpad, v_unpad, kv_lengths = prepare_data(
-        bsz, num_attn_heads, num_kv_heads, HEAD_DIM, same_context_len, Q_LEN, max_seq_len, dtype, device
-    )
-    max_seq_len_in_b = kv_lengths.max().item()  # for random lengths
-
-    quantiles = [0.5, 0.2, 0.8]
-    if provider == "torch":
-        k_torch = convert_kv_unpad_to_padded(k_unpad, kv_lengths, bsz, max_seq_len_in_b)
-        v_torch = convert_kv_unpad_to_padded(v_unpad, kv_lengths, bsz, max_seq_len_in_b)
-        torch_padding_mask = prepare_padding_mask(kv_lengths, bsz, max_seq_len_in_b, q.device)
-        fn = lambda: torch_attn_ref(
-            q, k_torch, v_torch, torch_padding_mask, bsz, 1, max_seq_len_in_b, num_attn_heads, num_kv_heads, HEAD_DIM
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench(fn, warmup=WARM_UPS, rep=REPS, quantiles=quantiles)
-    if provider == "triton":
-        k_cache, v_cache, block_tables = generate_caches_and_block_tables_v2(
-            k_unpad, v_unpad, kv_lengths, bsz, max_num_blocks_per_seq, block_size, dtype, device
-        )
-        block_tables = block_tables.to(device=device)
-        # the maximum block length splitted on kv should be the kv cache block size
-        kv_max_split_num = (max_seq_len_in_b + block_size - 1) // block_size
-        output = torch.empty((bsz, num_attn_heads, HEAD_DIM), dtype=dtype, device=device)
-        mid_output = torch.empty(
-            size=(bsz, num_attn_heads, kv_max_split_num, HEAD_DIM), dtype=torch.float32, device=q.device
-        )
-        mid_output_lse = torch.empty(size=(bsz, num_attn_heads, kv_max_split_num), dtype=torch.float32, device=q.device)
-        sm_scale = 1.0 / (HEAD_DIM**0.5)
-        fn = lambda: flash_decoding_attention(
-            # Here we use q.squeeze(2) because we hide the q_len dimension (which is equivalent to 1),
-            # refer to attention forward in modeling.
-            q.squeeze(2),
-            k_cache,
-            v_cache,
-            kv_lengths,
-            block_tables,
-            block_size,
-            max_seq_len_in_b,
-            output,
-            mid_output,
-            mid_output_lse,
-            sm_scale=sm_scale,
-            kv_group_num=kv_group_num,
-        )  # [bsz, 1, num_heads, head_dim]
-        ms, min_ms, max_ms = triton.testing.do_bench(fn, warmup=WARM_UPS, rep=REPS, quantiles=quantiles)
-
-    return ms, min_ms, max_ms
-
-
 if __name__ == "__main__":
     test_flash_decoding(16, 32, 32, 16, 1, True)
-    # bench_kernel.run(save_path=".", print_data=True)

tests/test_infer/test_ops/triton/test_fused_rotary_embedding.py

@@ -1,69 +1,11 @@
 from copy import deepcopy
 
 import torch
-import triton
 
 from colossalai.kernel.triton.fused_rotary_embedding import fused_rotary_embedding
 from colossalai.kernel.triton.no_pad_rotary_embedding import rotary_embedding
 from colossalai.kernel.triton.rotary_cache_copy import get_xine_cache
 
-BATCH = 16
-configs = [
-    triton.testing.Benchmark(
-        x_names=["num_tokens"],
-        x_vals=[2**i for i in range(4, 12)],
-        line_arg="provider",
-        line_vals=["torch_rotary_emb_func", "triton_rotary_emb_func"],
-        line_names=["torch_rotary_emb_func", "triton_rotary_emb_func"],
-        styles=[("red", "-"), ("blue", "-")],
-        ylabel="ms",
-        plot_name=f"rotary_emb-batch-{BATCH}",
-        args={"num_kv_heads": 16},
-    )
-]
-
-
-def torch_rotary_emb(x, cos, sin):
-    seq_len, h, dim = x.shape
-    x0 = x[:, :, 0 : dim // 2]
-    x1 = x[:, :, dim // 2 : dim]
-    cos = cos.view((seq_len, 1, dim // 2))
-    sin = sin.view((seq_len, 1, dim // 2))
-    o0 = x0 * cos - x1 * sin
-    o1 = x0 * sin + x1 * cos
-    return torch.cat((o0, o1), dim=-1)
-
-
-@triton.testing.perf_report(configs)
-def benchmark_rotary_emb(
-    provider: str,
-    num_tokens: int,
-    num_kv_heads: int,
-):
-    warmup = 10
-    rep = 100
-
-    head_dim = 128
-    dtype = torch.float16
-    q_shape = (num_tokens, num_kv_heads, head_dim)
-    q = -2.3 + 0.5 * torch.randn(q_shape, dtype=dtype, device="cuda")
-    k_shape = (num_tokens, num_kv_heads, head_dim)
-    k = -2.3 + 0.5 * torch.randn(k_shape, dtype=dtype, device="cuda")
-    cos_shape = (4096, head_dim // 2)
-    cos = -1.2 + 0.5 * torch.randn(cos_shape, dtype=dtype, device="cuda")
-    sin = -2.0 + 0.5 * torch.randn(cos_shape, dtype=dtype, device="cuda")
-
-    if provider == "torch_rotary_emb_func":
-        fn = lambda: torch_rotary_emb(q, cos[:num_tokens], sin[:num_tokens])
-    elif provider == "triton_rotary_emb_func":
-        fn = lambda: fused_rotary_embedding(q, k, cos, sin, lengths)
-    else:
-        raise ValueError("Undefined provider")
-
-    ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
-    return ms
-
-
 if __name__ == "__main__":
     num_tokens = 20
     num_kv_heads = 32
@@ -89,5 +31,3 @@ def benchmark_rotary_emb(
     fused_rotary_embedding(q_copy, k_copy, cos_cache, sin_cache, lengths)
     torch.allclose(q, q_copy)
     torch.allclose(k, k_copy)
-
-    # benchmark_rotary_emb.run(save_path=".",print_data=True)