biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc/Testing] Use torch.testing.assert_close (#7324)

Browse Source
This commit is contained in:
jon-chuang
2024-08-15 21:24:04 -07:00
committed by GitHub
parent e165528778
commit 50b8d08dbd
25 changed files with 197 additions and 188 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/kernels/quant_utils.py
View File

@@ -69,4 +69,4 @@ def ref_dynamic_per_tensor_fp8_quant(x: torch.tensor) \
ref_iscale = one / ref_scale
ref_out = (as_float32_tensor(x) * ref_iscale).clamp(
fp8_traits.min, fp8_traits.max).to(dtype=torch.float8_e4m3fn)
return ref_out, ref_scale
return ref_out, ref_scale.view((1, ))

10
tests/kernels/test_activation.py
View File

@@ -47,7 +47,7 @@ def test_act_and_mul(
ref_out = layer.forward_native(x)
# The SiLU and GELU implementations are equivalent to the native PyTorch
# implementations, so we can do exact comparison.
assert torch.allclose(out, ref_out, atol=0.0, rtol=0.0)
torch.testing.assert_close(out, ref_out, atol=0.0, rtol=0.0)
@pytest.mark.parametrize("activation", [FastGELU, NewGELU])
@@ -73,7 +73,7 @@ def test_activation(
layer = activation()
out = layer(x)
ref_out = layer.forward_native(x)
assert torch.allclose(out,
ref_out,
atol=get_default_atol(out),
rtol=get_default_rtol(out))
torch.testing.assert_close(out,
ref_out,
atol=get_default_atol(out),
rtol=get_default_rtol(out))

4
tests/kernels/test_attention.py
View File

@@ -276,7 +276,7 @@ def test_paged_attention(
atol, rtol = 1e-3, 1e-5
if kv_cache_dtype == "fp8":
atol, rtol = 1e-2, 1e-5
assert torch.allclose(output, ref_output, atol=atol, rtol=rtol)
torch.testing.assert_close(output, ref_output, atol=atol, rtol=rtol)
def ref_multi_query_kv_attention(
@@ -379,4 +379,4 @@ def test_multi_query_kv_attention(
)
atol = get_default_atol(output) if is_hip() else 1e-3
rtol = get_default_rtol(output) if is_hip() else 1e-5
assert torch.allclose(output, ref_output, atol=atol, rtol=rtol)
torch.testing.assert_close(output, ref_output, atol=atol, rtol=rtol)

4
tests/kernels/test_blocksparse_attention.py
View File

@@ -327,7 +327,7 @@ def test_paged_attention(
atol, rtol = 1e-3, 1e-5
if kv_cache_dtype == "fp8":
atol, rtol = 1e-2, 1e-5
assert torch.allclose(output, ref_output, atol=atol, rtol=rtol)
torch.testing.assert_close(output, ref_output, atol=atol, rtol=rtol)
def ref_multi_query_kv_attention(
@@ -441,4 +441,4 @@ def test_varlen_blocksparse_attention_prefill(
scale,
dtype,
)
assert torch.allclose(output, ref_output, atol=1e-2, rtol=1e-2)
torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2)

54
tests/kernels/test_cache.py
View File

@@ -98,10 +98,10 @@ def test_copy_blocks(
# Compare the results.
for key_cache, cloned_key_cache in zip(key_caches, cloned_key_caches):
assert torch.allclose(key_cache, cloned_key_cache)
torch.testing.assert_close(key_cache, cloned_key_cache)
for value_cache, cloned_value_cache in zip(value_caches,
cloned_value_caches):
assert torch.allclose(value_cache, cloned_value_cache)
torch.testing.assert_close(value_cache, cloned_value_cache)
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@@ -184,17 +184,17 @@ def test_reshape_and_cache(
cloned_value_cache[block_idx, :, :, block_offset] = value[i]
if kv_cache_dtype == "fp8":
assert torch.allclose(result_key_cache,
cloned_key_cache,
atol=0.001,
rtol=0.1)
assert torch.allclose(result_value_cache,
cloned_value_cache,
atol=0.001,
rtol=0.1)
torch.testing.assert_close(result_key_cache,
cloned_key_cache,
atol=0.001,
rtol=0.1)
torch.testing.assert_close(result_value_cache,
cloned_value_cache,
atol=0.001,
rtol=0.1)
else:
assert torch.allclose(key_cache, cloned_key_cache)
assert torch.allclose(value_cache, cloned_value_cache)
torch.testing.assert_close(key_cache, cloned_key_cache)
torch.testing.assert_close(value_cache, cloned_value_cache)
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@@ -290,17 +290,17 @@ def test_reshape_and_cache_flash(
cloned_value_cache[block_idx, block_offset, :, :] = value[i]
if kv_cache_dtype == "fp8":
assert torch.allclose(result_key_cache,
cloned_key_cache,
atol=0.001,
rtol=0.1)
assert torch.allclose(result_value_cache,
cloned_value_cache,
atol=0.001,
rtol=0.1)
torch.testing.assert_close(result_key_cache,
cloned_key_cache,
atol=0.001,
rtol=0.1)
torch.testing.assert_close(result_value_cache,
cloned_value_cache,
atol=0.001,
rtol=0.1)
else:
assert torch.allclose(key_cache, cloned_key_cache)
assert torch.allclose(value_cache, cloned_value_cache)
torch.testing.assert_close(key_cache, cloned_key_cache)
torch.testing.assert_close(value_cache, cloned_value_cache)
@pytest.mark.parametrize("direction", COPYING_DIRECTION)
@@ -372,10 +372,10 @@ def test_swap_blocks(
block_mapping_tensor)
for src, dst in block_mapping:
assert torch.allclose(src_key_caches_clone[src].cpu(),
dist_key_caches[0][dst].cpu())
assert torch.allclose(src_value_caches_clone[src].cpu(),
dist_value_caches[0][dst].cpu())
torch.testing.assert_close(src_key_caches_clone[src].cpu(),
dist_key_caches[0][dst].cpu())
torch.testing.assert_close(src_value_caches_clone[src].cpu(),
dist_value_caches[0][dst].cpu())
@pytest.mark.parametrize("num_heads", NUM_HEADS)
@@ -411,4 +411,4 @@ def test_fp8_e4m3_conversion(
converted_cache = torch.empty_like(cache)
ops.convert_fp8(converted_cache, cache_fp8)
assert torch.allclose(cache, converted_cache, atol=0.001, rtol=0.1)
torch.testing.assert_close(cache, converted_cache, atol=0.001, rtol=0.1)

23
tests/kernels/test_cutlass.py
View File

@@ -74,7 +74,7 @@ def cutlass_fp8_gemm_helper(m: int,
out = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
baseline = baseline_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
assert torch.allclose(out, baseline, rtol=1e-2, atol=5e-2)
torch.testing.assert_close(out, baseline, rtol=1e-2, atol=5e-2)
def cutlass_int8_gemm_helper(m: int,
@@ -106,7 +106,7 @@ def cutlass_int8_gemm_helper(m: int,
out = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
baseline = baseline_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
assert torch.allclose(out, baseline, rtol=1e-1, atol=1e0)
torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)
@pytest.mark.parametrize("m", [1, 16, 32, 64, 128, 256, 512, 222, 100, 33])
@@ -252,7 +252,7 @@ def test_cutlass_int8_azp_bias_fold(m: int, n: int, k: int,
azp_a = azp_aq_i8.to(dtype=torch.float32) * scale_a # correct for rounding
a_dq = scale_a * (aq_i32 + azp_aq_i8).to(dtype=torch.float32)
assert torch.allclose(a_dq, scale_a * aq_f32 + azp_a)
torch.testing.assert_close(a_dq, scale_a * aq_f32 + azp_a)
baseline_dq = torch.mm(a_dq, b_dq).to(out_dtype)
@@ -271,8 +271,8 @@ def test_cutlass_int8_azp_bias_fold(m: int, n: int, k: int,
scale_b,
out_dtype=out_dtype,
bias=azp_bias[0, :])
assert torch.allclose(out, baseline_dq, rtol=1e-2, atol=1e0)
assert torch.allclose(out, baseline_q, rtol=1e-2, atol=1e0)
torch.testing.assert_close(out, baseline_dq, rtol=1e-2, atol=1e0)
torch.testing.assert_close(out, baseline_q, rtol=1e-2, atol=1e0)
@pytest.mark.parametrize("m", [32, 64, 128])
@@ -302,7 +302,10 @@ def test_cutlass_int8_azp(m: int, n: int, k: int, out_dtype: torch.dtype,
azp_a = azp_aq_i8.to(dtype=torch.float32) * scale_a # correct for rounding
a_dq = scale_a * (aq_i32 - azp_aq_i8).to(dtype=torch.float32)
assert torch.allclose(a_dq, scale_a * aq_f32 - azp_a, rtol=1e-4, atol=1e-3)
torch.testing.assert_close(a_dq,
scale_a * aq_f32 - azp_a,
rtol=1e-4,
atol=1e-3)
if use_bias:
bias = torch.rand((1, n), device="cuda", dtype=out_dtype) * 10 + 2.5
@@ -335,8 +338,8 @@ def test_cutlass_int8_azp(m: int, n: int, k: int, out_dtype: torch.dtype,
# float16 precision is 10-bit mantissa -> 2^-11 ~ 0.05%
rtol = 1e-2 if out_dtype == torch.bfloat16 else 1e-3
atol = 1e-3
assert torch.allclose(out, baseline_dq, rtol=rtol, atol=atol)
assert torch.allclose(out, baseline_q, rtol=rtol, atol=atol)
torch.testing.assert_close(out, baseline_dq, rtol=rtol, atol=atol)
torch.testing.assert_close(out, baseline_q, rtol=rtol, atol=atol)
# Test working with a subset of A and B
@@ -363,7 +366,7 @@ def test_cutlass_subset():
scale_b,
out_dtype=torch.bfloat16)
assert torch.allclose(out, baseline, rtol=1e-1, atol=1e0)
torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)
# Test to make sure cuda graphs work
@@ -411,4 +414,4 @@ def test_cutlass_cuda_graph(per_act_token: bool, per_out_ch: bool):
baseline = torch.mm(scale_a * a.to(dtype=torch.float32),
scale_b * b.to(dtype=torch.float32)).to(torch.bfloat16)
assert torch.allclose(out, baseline, rtol=1e-1, atol=1e0)
torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)

4
tests/kernels/test_flash_attn.py
View File

@@ -126,7 +126,7 @@ def test_flash_attn_with_paged_kv(
scale=scale,
soft_cap=soft_cap,
)
assert torch.allclose(output, ref_output, atol=2e-2, rtol=1e-2), \
torch.testing.assert_close(output, ref_output, atol=2e-2, rtol=1e-2), \
f"{torch.max(torch.abs(output - ref_output))}"
@@ -211,5 +211,5 @@ def test_varlen_with_paged_kv(
sliding_window=sliding_window,
soft_cap=soft_cap,
)
assert torch.allclose(output, ref_output, atol=2e-2, rtol=1e-2), \
torch.testing.assert_close(output, ref_output, atol=2e-2, rtol=1e-2), \
f"{torch.max(torch.abs(output - ref_output))}"

4
tests/kernels/test_flashinfer.py
View File

@@ -144,7 +144,7 @@ def test_flashinfer_decode_with_paged_kv(kv_lens: List[int],
block_tables=block_tables,
scale=scale,
soft_cap=soft_cap)
assert torch.allclose(output, ref_output, atol=1e-2, rtol=1e-2), \
torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
f"{torch.max(torch.abs(output - ref_output))}"
@@ -244,5 +244,5 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
block_tables=block_tables,
scale=scale,
soft_cap=soft_cap)
assert torch.allclose(output, ref_output, atol=1e-2, rtol=1e-2), \
torch.testing.assert_close(output, ref_output, atol=1e-2, rtol=1e-2), \
f"{torch.max(torch.abs(output - ref_output))}"

14
tests/kernels/test_fp8_quant.py
View File

@@ -37,9 +37,9 @@ def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
scale_ub=scale_ub,
use_per_token_if_dynamic=True)
assert torch.allclose(ref_scales, ops_scales)
assert torch.allclose(ref_out.to(dtype=torch.float32),
ops_out.to(dtype=torch.float32))
torch.testing.assert_close(ref_scales, ops_scales)
torch.testing.assert_close(ref_out.to(dtype=torch.float32),
ops_out.to(dtype=torch.float32))
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@@ -57,9 +57,9 @@ def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
ref_out, ref_scale = ref_dynamic_per_tensor_fp8_quant(x)
ops_out, ops_scale = ops.scaled_fp8_quant(x)
assert torch.allclose(ref_scale, ops_scale)
assert torch.allclose(ref_out.to(dtype=torch.float32),
ops_out.to(dtype=torch.float32))
torch.testing.assert_close(ref_scale, ops_scale)
torch.testing.assert_close(ref_out.to(dtype=torch.float32),
ops_out.to(dtype=torch.float32))
# Regression test for a case with large activations where an int32 index cannot
@@ -84,4 +84,4 @@ def test_fp8_quant_large(seed: int) -> None:
ref_out = ref_out.to(dtype=dtype)
ops_out = ops_out.to(dtype=dtype)
assert torch.allclose(ref_out, ops_out)
torch.testing.assert_close(ref_out, ops_out)

12
tests/kernels/test_int8_quant.py
View File

@@ -29,9 +29,10 @@ def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
# kernel
ops_out, ops_scales = scaled_int8_quant(x)
assert torch.allclose(ops_scales, ref_scales)
assert torch.allclose(ops_out, ref_out,
atol=1) # big atol to account for rounding errors
torch.testing.assert_close(ops_scales, ref_scales)
torch.testing.assert_close(
ops_out, ref_out, atol=1,
rtol=0.0) # big atol to account for rounding errors
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@@ -54,5 +55,6 @@ def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
int8_traits.max).to(torch.int8)
out2, _ = scaled_int8_quant(x, scale)
assert torch.allclose(out1, out2,
atol=1) # big atol to account for rounding errors
torch.testing.assert_close(
out1, out2, atol=1,
rtol=0.0) # big atol to account for rounding errors

6
tests/kernels/test_layernorm.py
View File

@@ -48,7 +48,7 @@ def test_rms_norm(
# numerical errors than other operators because they involve reductions.
# Therefore, we use a larger tolerance.
if add_residual:
assert torch.allclose(out[0], ref_out[0], atol=1e-2, rtol=1e-2)
assert torch.allclose(out[1], ref_out[1], atol=1e-2, rtol=1e-2)
torch.testing.assert_close(out[0], ref_out[0], atol=1e-2, rtol=1e-2)
torch.testing.assert_close(out[1], ref_out[1], atol=1e-2, rtol=1e-2)
else:
assert torch.allclose(out, ref_out, atol=1e-2, rtol=1e-2)
torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2)

4
tests/kernels/test_marlin_gemm.py
View File

@@ -122,7 +122,7 @@ def test_gptq_marlin_repack(k_chunk, n_chunk, quant_type, group_size,
)
torch.cuda.synchronize()
assert torch.allclose(marlin_q_w_1, marlin_q_w_2)
torch.testing.assert_close(marlin_q_w_1, marlin_q_w_2)
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
@@ -174,7 +174,7 @@ def test_awq_marlin_repack(k_chunk, n_chunk, quant_type, group_size,
)
torch.cuda.synchronize()
assert torch.allclose(marlin_q_w_1, marlin_q_w_2)
torch.testing.assert_close(marlin_q_w_1, marlin_q_w_2)
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),

10
tests/kernels/test_moe.py
View File

@@ -50,7 +50,7 @@ def test_fused_moe(
score = torch.randn((m, e), device='cuda', dtype=dtype)
triton_output = fused_moe(a, w1, w2, score, topk, renormalize=False)
torch_output = torch_moe(a, w1, w2, score, topk)
assert torch.allclose(triton_output, torch_output, atol=1e-2, rtol=0)
torch.testing.assert_close(triton_output, torch_output, atol=1e-2, rtol=0)
@pytest.mark.parametrize("dtype",
@@ -95,7 +95,7 @@ def test_mixtral_moe(dtype: torch.dtype):
torch.bfloat16: 1e-2,
}
assert torch.allclose(hf_states.flatten(0, 1),
vllm_states,
rtol=mixtral_moe_tol[dtype],
atol=mixtral_moe_tol[dtype])
torch.testing.assert_close(hf_states.flatten(0, 1),
vllm_states,
rtol=mixtral_moe_tol[dtype],
atol=mixtral_moe_tol[dtype])

48
tests/kernels/test_pos_encoding.py
View File

@@ -67,14 +67,14 @@ def test_rotary_embedding(
ref_query, ref_key = rope.forward_native(positions, query, key)
out_query, out_key = rope.forward(positions, query, key)
# Compare the results.
assert torch.allclose(out_query,
ref_query,
atol=get_default_atol(out_query),
rtol=get_default_rtol(out_query))
assert torch.allclose(out_key,
ref_key,
atol=get_default_atol(out_key),
rtol=get_default_rtol(out_key))
torch.testing.assert_close(out_query,
ref_query,
atol=get_default_atol(out_query),
rtol=get_default_rtol(out_query))
torch.testing.assert_close(out_key,
ref_key,
atol=get_default_atol(out_key),
rtol=get_default_rtol(out_key))
@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
@@ -129,14 +129,14 @@ def test_batched_rotary_embedding(
dtype=torch.long,
device=device))
# Compare the results.
assert torch.allclose(out_query,
ref_query,
atol=get_default_atol(out_query),
rtol=get_default_rtol(out_query))
assert torch.allclose(out_key,
ref_key,
atol=get_default_atol(out_key),
rtol=get_default_rtol(out_key))
torch.testing.assert_close(out_query,
ref_query,
atol=get_default_atol(out_query),
rtol=get_default_rtol(out_query))
torch.testing.assert_close(out_key,
ref_key,
atol=get_default_atol(out_key),
rtol=get_default_rtol(out_key))
@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
@@ -200,14 +200,14 @@ def test_batched_rotary_embedding_multi_lora(
out_query, out_key = rope.forward(positions, query, key,
query_offsets.flatten())
# Compare the results.
assert torch.allclose(out_query,
ref_query,
atol=get_default_atol(out_query),
rtol=get_default_rtol(out_query))
assert torch.allclose(out_key,
ref_key,
atol=get_default_atol(out_key),
rtol=get_default_rtol(out_key))
torch.testing.assert_close(out_query,
ref_query,
atol=get_default_atol(out_query),
rtol=get_default_rtol(out_query))
torch.testing.assert_close(out_key,
ref_key,
atol=get_default_atol(out_key),
rtol=get_default_rtol(out_key))
@torch.inference_mode()

8
tests/kernels/test_sampler.py
View File

@@ -100,11 +100,11 @@ def test_sample_decoding_only(random_sampling, max_best_of,
if modify_greedy_probs and not request_uses_random_sampling:
# If we are modifying greedy probs and the request is greedy,
# we want to make sure the probs tensor is modified in place
assert torch.allclose(
torch.testing.assert_close(
probs[i][sampled_tokens[i]],
torch.full_like(probs[i][sampled_tokens[i]], 1.0))
assert torch.sum(probs[i]) == 1.0
assert torch.allclose(
torch.testing.assert_close(
sampled_modified_probs[i][0],
torch.full_like(sampled_modified_probs[i][0], 1.0))
elif request_uses_random_sampling:
@@ -117,8 +117,8 @@ def test_sample_decoding_only(random_sampling, max_best_of,
# If the request is greedy and we are not modifying greedy probs,
# we want to make sure sampled_modified_probs tensor is the same as
# the probs tensor.
assert torch.allclose(sampled_modified_probs[i][0],
probs[i][sampled_tokens[i]])
torch.testing.assert_close(sampled_modified_probs[i],
probs[i][sampled_tokens[i]])
if save_logprobs:
assert sampled_logprobs.shape == (bs, max_best_of)

4
tests/kernels/utils.py
View File

@@ -924,5 +924,5 @@ def assert_actual_matches_ideal(test_params: PhaseTestParameters,
* output_under_test: actually observed output value
'''
ideal_output = test_params.packed_qkvo.ideal_output
assert torch.allclose(ideal_output,
output_under_test.view_as(ideal_output))
torch.testing.assert_close(ideal_output,
output_under_test.view_as(ideal_output))