[Kimi-K3] DCP support (#50484)

Signed-off-by: girasoley <girasoley@inferact.ai>
Signed-off-by: Summer Yang <girasoleyang@gmail.com>
Signed-off-by: 云挚 <ningyunxiao.nyx@antgroup.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: 云挚 <ningyunxiao.nyx@antgroup.com>
Co-authored-by: foraxe <73625538+foraxe@users.noreply.github.com>
Co-authored-by: Woosuk Kwon <woosuk@inferact.ai>
This commit is contained in:
Summer Yang
2026-08-10 12:54:26 -07:00
committed by GitHub
co-authored by Claude Fable 5 云挚 foraxe Woosuk Kwon
parent 3f142bd85e
commit 63ac04a61e
19 changed files with 3529 additions and 82 deletions
+106 -2
View File
@@ -222,6 +222,49 @@ class TestLSEWeightedCombine:
assert result.shape == (B, H, D)
torch.testing.assert_close(result, outputs[1], atol=1e-5, rtol=1e-5)
def test_empty_shard_ignores_undefined_output(self):
from vllm.v1.attention.ops.dcp_alltoall import _lse_weighted_combine
outputs = torch.tensor([[[[float("nan")]]], [[[3.0]]]])
lses = torch.tensor([[[-float("inf")]], [[0.0]]])
result = _lse_weighted_combine(outputs, lses)
torch.testing.assert_close(result, outputs[1])
def test_ag_rs_masks_empty_shard_and_padded_lse(self, monkeypatch):
import vllm.v1.attention.ops.common as common
class FakeGroup:
world_size = 2
rank_in_group = 0
def all_gather(self, tensor, dim):
assert dim == 0
return torch.cat((tensor, tensor), dim=dim)
monkeypatch.setattr(
common,
"correct_attn_out",
lambda output, lses, *args, **kwargs: (output, lses[0]),
)
output = torch.ones(7, 1, 1)
lse = torch.ones(7, 1)
seq_lens = torch.tensor([0, 2], dtype=torch.int32)
query_start_loc = torch.tensor([0, 1, 5], dtype=torch.int32)
_, masked_lse = common._cp_lse_common(
output,
lse,
FakeGroup(),
seq_lens=seq_lens,
query_start_loc=query_start_loc,
)
assert torch.isneginf(masked_lse[:1]).all()
torch.testing.assert_close(masked_lse[1:5], torch.ones_like(masked_lse[1:5]))
assert torch.isneginf(masked_lse[5:]).all()
def test_mathematically_correct(self):
"""Verify mathematical correctness of LSE combination."""
from vllm.v1.attention.ops.dcp_alltoall import _lse_weighted_combine
@@ -344,7 +387,7 @@ class TestPackedA2AKernels:
world_size, B, h_per_rank, D = 4, 7, 2, 32
H = world_size * h_per_rank
cp_attn_out = torch.randn(B, H, D, device=device, dtype=dtype)
cp_attn_lse = torch.randn(B, H, device=device, dtype=torch.float32)
cp_attn_lse = torch.randn(B, H, device=device, dtype=dtype)
lse_pack_dim = _dcp_a2a_lse_pack_dim(dtype)
send_buffer = torch.empty(
(world_size, B, h_per_rank, D + lse_pack_dim),
@@ -375,6 +418,67 @@ class TestPackedA2AKernels:
else:
_assert_packed_a2a_close(actual, expected_out, dtype)
@pytest.mark.skipif(
torch.accelerator.device_count() < 1, reason="CUDA is required."
)
def test_empty_seq_lens_ignore_undefined_output(self):
from vllm.v1.attention.ops.dcp_alltoall import (
_dcp_a2a_lse_pack_dim,
_dcp_a2a_pack_send,
_dcp_a2a_unpack_combine,
)
device = torch.device("cuda")
world_size, num_tokens, h_per_rank, head_dim = 2, 5, 1, 32
num_heads = world_size * h_per_rank
output = torch.randn(
num_tokens,
num_heads,
head_dim,
device=device,
dtype=torch.bfloat16,
)
output[:1] = float("nan")
lse = torch.randn(num_tokens, num_heads, device=device, dtype=output.dtype)
seq_lens = torch.tensor([0, 2], device=device, dtype=torch.int32)
query_start_loc = torch.tensor([0, 1, 5], device=device, dtype=torch.int32)
lse_pack_dim = _dcp_a2a_lse_pack_dim(output.dtype)
send_buffer = torch.empty(
(
world_size,
num_tokens,
h_per_rank,
head_dim + lse_pack_dim,
),
device=device,
dtype=output.dtype,
)
_dcp_a2a_pack_send(
output,
lse,
send_buffer,
world_size,
h_per_rank,
head_dim,
lse_pack_dim,
seq_lens=seq_lens,
query_start_loc=query_start_loc,
)
actual_output, actual_lse = _dcp_a2a_unpack_combine(
send_buffer,
head_dim,
lse_pack_dim,
return_lse=True,
is_lse_base_on_e=True,
)
torch.testing.assert_close(
actual_output[:1], torch.zeros_like(actual_output[:1])
)
assert torch.isneginf(actual_lse[:1]).all()
assert torch.isfinite(actual_output[1:]).all()
def _distributed_packed_a2a_worker(env: dict[str, str]) -> None:
update_environment_variables(env)
@@ -417,7 +521,7 @@ def _distributed_packed_a2a_worker(env: dict[str, str]) -> None:
B,
H,
device=f"cuda:{local_rank}",
dtype=torch.float32,
dtype=dtype,
generator=generator,
)
actual = dcp_a2a_lse_reduce(
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,128 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import math
import pytest
from tests.conftest import VllmRunner
from tests.utils import create_new_process_for_each_test
from vllm import SamplingParams, TokensPrompt
MODEL = "moonshotai/Kimi-Linear-48B-A3B-Instruct"
def _make_tiny_overrides() -> dict:
linear_attn_config = {
"full_attn_layers": [4],
"head_dim": 128,
"kda_layers": [1, 2, 3],
"num_heads": 8,
"short_conv_kernel_size": 4,
"use_full_rank_gate": True,
}
return {
"bos_token_id": 1,
"eos_token_id": 2,
"pad_token_id": 0,
"vocab_size": 256,
"hidden_size": 256,
"head_dim": 32,
"intermediate_size": 512,
"num_hidden_layers": 4,
"num_attention_heads": 8,
"num_key_value_heads": 8,
"num_experts": None,
"num_experts_per_token": None,
"num_shared_experts": 0,
"q_lora_rank": 128,
"kv_lora_rank": 512,
"qk_nope_head_dim": 128,
"qk_rope_head_dim": 64,
"v_head_dim": 128,
"mla_use_nope": True,
"mla_use_output_gate": True,
"linear_attn_config": linear_attn_config,
"max_position_embeddings": 512,
"model_max_length": 512,
}
def _run_tiny_model(
vllm_runner: type[VllmRunner], dcp_size: int
) -> list[tuple[list[int], list[float]]]:
with vllm_runner(
model_name=MODEL,
skip_tokenizer_init=True,
load_format="dummy",
hf_overrides=_make_tiny_overrides(),
tensor_parallel_size=2,
decode_context_parallel_size=dcp_size,
cp_kv_cache_interleave_size=1,
distributed_executor_backend="mp",
dtype="bfloat16",
seed=0,
enforce_eager=True,
max_model_len=512,
max_num_seqs=8,
max_num_batched_tokens=1024,
gpu_memory_utilization=0.25,
enable_chunked_prefill=True,
enable_prefix_caching=False,
) as runner:
lengths = [1, 32, 33, 129]
prompts = [
TokensPrompt(
prompt_token_ids=[
3 + ((request_idx + token_idx) % 251) for token_idx in range(length)
]
)
for request_idx, length in enumerate(lengths)
]
outputs = runner.llm.generate(
prompts,
SamplingParams(
temperature=0,
max_tokens=8,
seed=0,
ignore_eos=True,
logprobs=20,
),
use_tqdm=False,
)
results = []
for request_output in outputs:
completion = request_output.outputs[0]
token_ids = list(completion.token_ids)
assert completion.logprobs is not None
selected_logprobs = [
step_logprobs[token_id].logprob
for token_id, step_logprobs in zip(token_ids, completion.logprobs)
]
results.append((token_ids, selected_logprobs))
return results
@create_new_process_for_each_test()
@pytest.mark.distributed(num_gpus=2)
def test_kimi_linear_dcp_tiny(
vllm_runner: type[VllmRunner],
num_gpus_available: int,
) -> None:
if num_gpus_available < 2:
pytest.skip("Need at least 2 GPUs")
baseline = _run_tiny_model(vllm_runner, dcp_size=1)
dcp = _run_tiny_model(vllm_runner, dcp_size=2)
assert [tokens for tokens, _ in dcp] == [tokens for tokens, _ in baseline]
logprob_drifts = []
for (_, baseline_logprobs), (_, dcp_logprobs) in zip(baseline, dcp):
for baseline_logprob, dcp_logprob in zip(baseline_logprobs, dcp_logprobs):
assert math.isfinite(baseline_logprob)
assert math.isfinite(dcp_logprob)
logprob_drifts.append(abs(baseline_logprob - dcp_logprob))
assert max(logprob_drifts) <= 1e-2
@@ -0,0 +1,246 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
import vllm._custom_ops as ops
from vllm.models.kimi_k3.nvidia.ops.fused_mla_key_concat_kv_cache import (
fused_mla_decode_q_concat_kv_cache_insert,
fused_mla_key_concat_ds_mla_insert,
fused_mla_key_concat_kv_cache_insert,
fused_mla_qkv_quant_kv_cache_fp8_insert,
)
from vllm.platforms import current_platform
NUM_TOKENS = 4
NUM_HEADS = 2
KV_LORA_RANK = 512
NOPE_HEAD_DIM = 128
ROPE_HEAD_DIM = 64
QK_HEAD_DIM = NOPE_HEAD_DIM + ROPE_HEAD_DIM
V_HEAD_DIM = 128
CACHE_ENTRY = KV_LORA_RANK + ROPE_HEAD_DIM
DS_MLA_CACHE_ENTRY = 656
BLOCK_SIZE = 8
OWNED_TOKEN_INDICES = [1, 3]
pytestmark = pytest.mark.skipif(
not current_platform.is_cuda(), reason="Kimi-K3 fused MLA ops require CUDA"
)
def _inputs() -> dict[str, torch.Tensor]:
torch.manual_seed(0)
def make(*shape: int) -> torch.Tensor:
return torch.randn(*shape, device="cuda", dtype=torch.bfloat16)
return {
"q": make(NUM_TOKENS, NUM_HEADS, QK_HEAD_DIM),
"k_nope": make(NUM_TOKENS, NUM_HEADS, NOPE_HEAD_DIM),
"k_pe": make(NUM_TOKENS, ROPE_HEAD_DIM),
"kv_c": make(NUM_TOKENS, KV_LORA_RANK),
"v": make(NUM_TOKENS, NUM_HEADS, V_HEAD_DIM),
"ql_nope": make(NUM_TOKENS, NUM_HEADS, KV_LORA_RANK),
"q_pe": make(NUM_TOKENS, NUM_HEADS, ROPE_HEAD_DIM),
}
def _slot_mapping() -> torch.Tensor:
return torch.tensor([-1, 0, -1, 5], device="cuda", dtype=torch.int64)
def _owned(tensor: torch.Tensor) -> torch.Tensor:
return tensor[OWNED_TOKEN_INDICES]
def _full_key(k_nope: torch.Tensor, k_pe: torch.Tensor) -> torch.Tensor:
shared_pe = k_pe[:, None, :].expand(-1, k_nope.shape[1], -1)
return torch.cat((k_nope, shared_pe), dim=-1)
def _latent_query(ql_nope: torch.Tensor, q_pe: torch.Tensor) -> torch.Tensor:
return torch.cat((ql_nope, q_pe), dim=-1)
def _make_cache(cache_format: str) -> torch.Tensor:
if cache_format == "bf16":
return torch.full(
(1, BLOCK_SIZE, CACHE_ENTRY),
-7,
device="cuda",
dtype=torch.bfloat16,
)
if cache_format == "fp8":
return torch.full(
(1, BLOCK_SIZE, CACHE_ENTRY),
1.0,
device="cuda",
dtype=torch.float32,
).to(torch.float8_e4m3fn)
assert cache_format == "fp8_ds_mla"
return torch.full(
(1, BLOCK_SIZE, DS_MLA_CACHE_ENTRY),
165,
device="cuda",
dtype=torch.uint8,
)
def _reference_cache(
cache_format: str,
inputs: dict[str, torch.Tensor],
initial_cache: torch.Tensor,
slots: torch.Tensor,
) -> torch.Tensor:
reference_cache = initial_cache.clone()
valid_slots = _owned(slots)
if cache_format == "fp8_ds_mla":
ops.concat_and_cache_mla(
_owned(inputs["kv_c"]),
_owned(inputs["k_pe"]),
reference_cache,
valid_slots,
cache_format,
torch.ones(1, device="cuda", dtype=torch.float32),
)
return reference_cache
latent = torch.cat(
(_owned(inputs["kv_c"]), _owned(inputs["k_pe"])),
dim=-1,
)
flat_reference_cache = reference_cache.view(-1, CACHE_ENTRY)
for row, slot in zip(latent, valid_slots.tolist()):
flat_reference_cache[slot].copy_(row.to(reference_cache.dtype))
return reference_cache
def _assert_cache_matches_reference(
cache: torch.Tensor,
reference_cache: torch.Tensor,
initial_cache: torch.Tensor,
) -> None:
assert torch.equal(cache, reference_cache)
for slot in (0, 5):
assert not torch.equal(cache[0, slot], initial_cache[0, slot])
for slot in (1, 2, 3, 4, 6, 7):
assert torch.equal(cache[0, slot], initial_cache[0, slot])
@pytest.mark.parametrize("cache_format", ["bf16", "fp8_ds_mla"])
def test_prefill_concat_ignores_negative_slots_for_cache(cache_format: str) -> None:
inputs = _inputs()
slots = _slot_mapping()
initial_cache = _make_cache(cache_format)
mixed_cache = initial_cache.clone()
reference_cache = _reference_cache(cache_format, inputs, initial_cache, slots)
op = (
fused_mla_key_concat_kv_cache_insert
if cache_format == "bf16"
else fused_mla_key_concat_ds_mla_insert
)
output = op(
inputs["q"],
inputs["k_nope"],
inputs["k_pe"],
inputs["kv_c"],
mixed_cache,
slots,
)
torch.testing.assert_close(
output, _full_key(inputs["k_nope"], inputs["k_pe"]), rtol=0, atol=0
)
_assert_cache_matches_reference(mixed_cache, reference_cache, initial_cache)
def test_fp8_prefill_outputs_ignore_negative_slots_for_cache() -> None:
inputs = _inputs()
slots = _slot_mapping()
initial_cache = _make_cache("fp8")
mixed_cache = initial_cache.clone()
reference_cache = _reference_cache("fp8", inputs, initial_cache, slots)
scale_inv = torch.ones(1, device="cuda", dtype=torch.float32)
q, k, v = fused_mla_qkv_quant_kv_cache_fp8_insert(
inputs["q"],
inputs["k_nope"],
inputs["k_pe"],
inputs["kv_c"],
inputs["v"],
mixed_cache,
slots,
scale_inv,
scale_inv,
scale_inv,
scale_inv,
)
expected = (
inputs["q"].to(torch.float8_e4m3fn),
_full_key(inputs["k_nope"], inputs["k_pe"]).to(torch.float8_e4m3fn),
inputs["v"].to(torch.float8_e4m3fn),
)
for actual, reference in zip((q, k, v), expected):
torch.testing.assert_close(actual.float(), reference.float(), rtol=0, atol=0)
_assert_cache_matches_reference(mixed_cache, reference_cache, initial_cache)
@pytest.mark.parametrize("cache_format", ["bf16", "fp8", "fp8_ds_mla"])
def test_decode_concat_ignores_negative_slots_for_cache(cache_format: str) -> None:
inputs = _inputs()
slots = _slot_mapping()
initial_cache = _make_cache(cache_format)
mixed_cache = initial_cache.clone()
reference_cache = _reference_cache(cache_format, inputs, initial_cache, slots)
scale_inv = torch.ones(1, device="cuda", dtype=torch.float32)
kwargs = {
"ds_mla": cache_format == "fp8_ds_mla",
"q_scale_inv": scale_inv if cache_format == "fp8" else None,
"cache_scale_inv": scale_inv if cache_format == "fp8" else None,
}
output = fused_mla_decode_q_concat_kv_cache_insert(
inputs["ql_nope"],
inputs["q_pe"],
inputs["kv_c"],
inputs["k_pe"],
mixed_cache,
slots,
**kwargs,
)
expected = _latent_query(inputs["ql_nope"], inputs["q_pe"])
if cache_format == "fp8":
expected = expected.to(torch.float8_e4m3fn)
torch.testing.assert_close(output.float(), expected.float(), rtol=0, atol=0)
else:
torch.testing.assert_close(output, expected, rtol=0, atol=0)
_assert_cache_matches_reference(mixed_cache, reference_cache, initial_cache)
def test_ds_mla_cache_insert_bit_compatible_with_reference() -> None:
"""Fused ds_mla insertion must match the reference bit-for-bit."""
torch.manual_seed(0)
num_tokens, num_blocks = 33, 16
dt = torch.bfloat16
kv_c = torch.randn(num_tokens, KV_LORA_RANK, device="cuda", dtype=dt)
k_pe = torch.randn(num_tokens, 1, ROPE_HEAD_DIM, device="cuda", dtype=dt)
k_nope = torch.randn(num_tokens, NUM_HEADS, NOPE_HEAD_DIM, device="cuda", dtype=dt)
q = torch.randn(num_tokens, NUM_HEADS, QK_HEAD_DIM, device="cuda", dtype=dt)
slots = torch.randperm(num_blocks * BLOCK_SIZE, device="cuda")[:num_tokens].long()
slots[torch.rand(num_tokens, device="cuda") < 0.3] = -1
ref = torch.full(
(num_blocks, BLOCK_SIZE, DS_MLA_CACHE_ENTRY),
165,
device="cuda",
dtype=torch.uint8,
)
got = ref.clone()
scale = torch.ones(1, device="cuda", dtype=torch.float32)
ops.concat_and_cache_mla(kv_c, k_pe.squeeze(1), ref, slots, "fp8_ds_mla", scale)
fused_mla_key_concat_ds_mla_insert(q, k_nope, k_pe, kv_c, got, slots)
assert torch.equal(ref, got)
@@ -0,0 +1,80 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from unittest.mock import MagicMock
import pytest
import torch
from vllm.platforms import current_platform
requires_flashinfer_mla = pytest.mark.skipif(
not current_platform.has_device_capability(100),
reason="FlashInfer MLA requires compute capability 10 or above.",
)
def test_mla_dcp_gathered_query_reserves_backend_head_storage():
from vllm.v1.attention.ops.dcp_utils import reserve_query_head_storage
query = torch.randn(3, 24, 576, dtype=torch.bfloat16)
padded = reserve_query_head_storage(query, 128)
assert padded.shape == query.shape
assert padded.stride() == query.stride()
assert padded.untyped_storage().nbytes() >= 3 * 128 * 576 * 2
assert torch.equal(padded, query)
@requires_flashinfer_mla
def test_flashinfer_mla_selects_backend_for_gathered_heads():
from vllm.v1.attention.backends.mla.flashinfer_mla import (
_select_mla_decode_backend,
)
assert _select_mla_decode_backend(6) is None
assert _select_mla_decode_backend(24) == "cute-dsl"
@requires_flashinfer_mla
def test_flashinfer_mla_forward_uses_gathered_head_count(monkeypatch):
import vllm.v1.attention.backends.mla.flashinfer_mla as flashinfer_mla
impl = MagicMock()
impl.bmm1_scale = 1.0
impl.bmm2_scale = 1.0
impl.need_to_return_lse_for_decode = True
impl.num_heads = 6
impl.qk_nope_head_dim = 128
impl.kv_lora_rank = 512
impl.qk_rope_head_dim = 64
impl.kv_cache_dtype = "auto"
attn_metadata = MagicMock()
attn_metadata.causal = True
attn_metadata.num_decode_tokens = 2
attn_metadata.num_decodes = 2
attn_metadata.max_seq_len = 1
attn_metadata.decode.block_table = torch.zeros(2, 1, dtype=torch.int32)
attn_metadata.decode.seq_lens = torch.ones(2, dtype=torch.int32)
query = torch.ones(2, 24, 576, dtype=torch.bfloat16)
kv_cache = torch.ones(1, 128, 576, dtype=torch.bfloat16)
kernel = MagicMock(
return_value=(
torch.ones(2, 1, 24, 512, dtype=torch.bfloat16),
torch.ones(2, 24, dtype=torch.float32),
)
)
monkeypatch.setattr(flashinfer_mla, "_get_workspace_buffer", MagicMock())
monkeypatch.setattr(flashinfer_mla, "trtllm_batch_decode_with_kv_cache_mla", kernel)
layer = MagicMock()
output, lse = flashinfer_mla.FlashInferMLAImpl.forward_mqa(
impl, query, kv_cache, attn_metadata, layer
)
assert kernel.call_args.kwargs["backend"] == "cute-dsl"
assert output.shape == (2, 24, 512)
assert lse is not None and lse.shape == (2, 24)