minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Fix nightly CI: NVFP4 cuda-graph crash, NVILA batching, CuTe paged-KV zero-size, Kimi-VL OOM (#32118)

Browse Source 
Co-authored-by: Brayden Zhong <brayden@radixark.ai>
This commit is contained in:
Brayden Zhong
2026-07-29 01:39:55 -07:00
committed by GitHub
co-authored by Brayden Zhong
parent cce5873513
commit 7dcebca255
5 changed files with 48 additions and 221 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/kernels/ops/attention/flash_attn/cute/interface.py
+6 -1
View File
@@ -692,7 +692,12 @@ def _flash_attn_fwd(
# SplitKV uses float32 partial output, which doubles the O buffer size
# in shared memory, causing OOM for diff-headdim (192, 128)
if arch // 10 in [10, 11] and head_dim != head_dim_v and num_splits > 1:
if (
arch // 10 in [10, 11]
and head_dim != head_dim_v
and num_splits > 1
and not (page_table is not None and q_stage == 1)
):
if num_n_blocks >= 64 and head_dim_v != 512:
tile_n = 64
num_n_blocks = (seqlen_k_loaded + tile_n - 1) // tile_n
python/sglang/srt/models/kimi_vl.py
+6 -1
View File
@@ -131,7 +131,6 @@ class KimiVLForConditionalGeneration(nn.Module):
config.vision_config,
prefix=add_prefix("vision_tower", prefix),
use_data_parallel=self.use_data_parallel,
use_tensor_parallel=not self.use_data_parallel,
)
self.multi_modal_projector = KimiVLMultiModalProjector(config=config)
@@ -263,6 +262,12 @@ class KimiVLForConditionalGeneration(nn.Module):
use_default_weight_loading = False
if "vision" in name:
if self.vision_tower is not None:
# MoonViT's attention is wrapped in sglang's VisionAttention,
# whose sub-modules are named qkv_proj/proj instead of the
# checkpoint's wqkv/wo.
name = name.replace("wqkv.", "attn.qkv_proj.").replace(
"wo.", "attn.proj."
)
use_default_weight_loading = True
else:
for param_name, weight_name, shard_id in stacked_params_mapping:
python/sglang/srt/models/kimi_vl_moonvit.py
+30 -198
View File
@@ -52,18 +52,15 @@ import torch.nn.functional as F
from transformers.activations import ACT2FN
from transformers.modeling_utils import PreTrainedModel
from sglang.kernel_api_logging import debug_kernel_api
try:
from flash_attn.flash_attn_interface import flash_attn_varlen_func
except ImportError:
flash_attn_varlen_func = None
from sglang.srt.configs import MoonViTConfig
from sglang.srt.layers.attention.vision import (
VisionAttention,
VisionAttentionMetadata,
prepare_vision_attention_metadata,
)
from sglang.srt.layers.conv import Conv2dLayer
from sglang.srt.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
ReplicatedLinear,
RowParallelLinear,
)
@@ -75,112 +72,6 @@ from sglang.srt.utils import add_prefix, get_device
_MAX_INFERENCE_POS_EMB_CACHE_ENTRIES = 256
@debug_kernel_api
def multihead_attention(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
q_cu_seqlens: Optional[torch.Tensor] = None,
k_cu_seqlens: Optional[torch.Tensor] = None,
max_seqlen: Optional[int] = None,
):
"""Multi-head attention using flash attention 2.
This function is used to handle the case where the query, key, and value are packed.
Args:
q, k, v: tensor of shape (tot_seqlens, num_heads, head_dim).
q_cu_seqlens (torch.Tensor): cumulative sequence lengths of q.
The first element should be 0 and the last element should be q.shape[0].
k_cu_seqlens (torch.Tensor): cumulative sequence lengths of k.
The first element should be 0 and the last element should be k.shape[0].
Returns:
output: shape (batch_size, seqlen, dim) or (tot_seqlens, dim) if packing,
where dim = num_heads * head_dim
"""
if flash_attn_varlen_func is None:
raise ImportError(
"flash_attn is not installed, this function needs flash_attn_varlen_func from flash_attn"
)
# Unified format legal check
assert q.dim() == k.dim() == v.dim() == 3, "q, k, v must have 3 dims"
# Keep validation on CPU for debugging, but avoid synchronizing the GPU
# once per MoonViT layer in the normal packed CUDA path.
if not q_cu_seqlens.is_cuda:
assert q_cu_seqlens[-1] == q.shape[0], "q_cu_seqlens must sum to q.shape[0]"
assert (
k_cu_seqlens[-1] == k.shape[0] == v.shape[0]
), "k_cu_seqlens must sum to k.shape[0]"
assert q.dtype in [
torch.bfloat16,
torch.float16,
], f"unsupported dtype {q.dtype} for multihead attn"
if max_seqlen is None:
max_seqlen = (q_cu_seqlens[1:] - q_cu_seqlens[:-1]).max().item()
attn_out = flash_attn_varlen_func(
q,
k,
v,
q_cu_seqlens,
k_cu_seqlens,
max_seqlen,
max_seqlen,
causal=False,
)
attn_out = attn_out.flatten(start_dim=-2)
return attn_out
def sdpa_attention(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
q_cu_seqlens: Optional[torch.Tensor] = None,
k_cu_seqlens: Optional[torch.Tensor] = None,
max_seqlen: Optional[int] = None,
) -> torch.Tensor:
"""Multi-head attention using torch scaled dot product attention.
This function is used to handle the case where the query, key, and value are packed.
Args:
q, k, v: tensor of shape (tot_seqlens, num_heads, head_dim).
q_cu_seqlens (torch.Tensor): cumulative sequence lengths of q.
The first element should be 0 and the last element should be q.shape[0].
k_cu_seqlens (torch.Tensor): cumulative sequence lengths of k.
The first element should be 0 and the last element should be k.shape[0].
Returns:
output: shape (batch_size, seqlen, dim) or (tot_seqlens, dim) if packing,
where dim = num_heads * head_dim
"""
# Unified format legal check
assert q.dim() == k.dim() == v.dim() == 3, "q, k, v must have 3 dims"
assert q_cu_seqlens[-1] == q.shape[0], "q_cu_seqlens must sum to q.shape[0]"
seq_length = q.shape[0]
attention_mask = torch.zeros(
[1, seq_length, seq_length], device=q.device, dtype=torch.bool
)
for i in range(1, len(q_cu_seqlens)):
attention_mask[
...,
q_cu_seqlens[i - 1] : q_cu_seqlens[i],
q_cu_seqlens[i - 1] : q_cu_seqlens[i],
] = True
q = q.transpose(0, 1)
k = k.transpose(0, 1)
v = v.transpose(0, 1)
attn_output = F.scaled_dot_product_attention(q, k, v, attention_mask, dropout_p=0.0)
attn_output = attn_output.transpose(0, 1)
attn_output = attn_output.reshape(seq_length, -1)
return attn_output
VL_VISION_ATTENTION_FUNCTIONS = {
"flash_attention_2": multihead_attention,
"sdpa": sdpa_attention,
}
def _apply_rope_input_validation(x, freqs_cis):
assert x.ndim == freqs_cis.ndim + 1, (x.shape, freqs_cis.shape)
assert x.shape[:-2] == freqs_cis.shape[:-1], (x.shape, freqs_cis.shape)
@@ -189,7 +80,7 @@ def _apply_rope_input_validation(x, freqs_cis):
def apply_rope(
xq: torch.Tensor, xk: torch.Tensor, freqs_cis: torch.Tensor
xq: torch.Tensor, xk: torch.Tensor, freqs_cis: torch.Tensor, x_shape=None
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Args: (The leading dimensions of all inputs should be the same)
@@ -518,24 +409,17 @@ class MoonVitEncoderLayer(nn.Module):
hidden_dim: int,
mlp_dim: int,
*,
attn_implementation: str = "flash_attention_2", # use fa2 in sglang by default
activation=F.gelu,
attn_bias: bool = False,
prefix: str = "",
use_data_parallel: bool = False,
use_tensor_parallel: bool = False,
):
super().__init__()
self.num_heads = num_heads
self.hidden_dim = hidden_dim
self.hidden_size_per_attention_head = self.hidden_dim // self.num_heads
self.attn_implementation = attn_implementation
self.use_tensor_parallel = use_tensor_parallel and not use_data_parallel
tp_size = get_parallel().attn_tp_size if self.use_tensor_parallel else 1
tp_rank = get_parallel().attn_tp_rank if self.use_tensor_parallel else 0
self.num_attention_heads_per_partition = self.num_heads // tp_size
use_tensor_parallel = not use_data_parallel
self.norm0 = nn.LayerNorm(hidden_dim)
self.norm1 = nn.LayerNorm(hidden_dim)
self.mlp = MLP2(
@@ -543,74 +427,20 @@ class MoonVitEncoderLayer(nn.Module):
activation,
prefix=add_prefix("mlp", prefix),
use_data_parallel=use_data_parallel,
use_tensor_parallel=self.use_tensor_parallel,
use_tensor_parallel=use_tensor_parallel,
)
if self.use_tensor_parallel:
self.wqkv = QKVParallelLinear(
hidden_size=hidden_dim,
head_size=self.hidden_size_per_attention_head,
total_num_heads=num_heads,
total_num_kv_heads=num_heads,
bias=attn_bias,
prefix=add_prefix("wqkv", prefix),
tp_rank=tp_rank,
tp_size=tp_size,
)
self.wo = RowParallelLinear(
hidden_dim,
hidden_dim,
bias=attn_bias,
prefix=add_prefix("wo", prefix),
tp_rank=tp_rank,
tp_size=tp_size,
)
else:
self.wqkv = nn.Linear(hidden_dim, hidden_dim * 3, bias=attn_bias)
self.wo = nn.Linear(hidden_dim, hidden_dim, bias=attn_bias)
def attention_qkvpacked(
self,
x: torch.Tensor,
cu_seqlens: torch.Tensor,
rope_freqs_cis: Optional[torch.Tensor] = None,
max_seqlen: Optional[int] = None,
):
"""
Args:
x (torch.Tensor): (batch_size, seqlen, hidden_dim)
cu_seqlens (torch.Tensor):
"""
if self.use_tensor_parallel:
xqkv, _ = self.wqkv(x)
else:
xqkv = self.wqkv(x)
qkv_shape = xqkv.size()[:-1] + (
3,
self.num_attention_heads_per_partition,
self.hidden_size_per_attention_head,
self.attn = VisionAttention(
embed_dim=hidden_dim,
num_heads=num_heads,
projection_size=hidden_dim,
use_qkv_parallel=True,
qkv_bias=attn_bias,
proj_bias=attn_bias,
flatten_batch=True,
prefix=add_prefix("attn", prefix),
use_data_parallel=use_data_parallel,
customized_position_embedding_applier=apply_rope,
)
# xqkv: (batch_size, seqlen, 3, nheads, headdim)
xqkv = xqkv.view(*qkv_shape)
xq, xk, xv = torch.unbind(xqkv, dim=-3)
xq, xk = apply_rope(xq, xk, rope_freqs_cis)
attn_func = VL_VISION_ATTENTION_FUNCTIONS[self.attn_implementation]
attn_out = attn_func(
xq,
xk,
xv,
q_cu_seqlens=cu_seqlens,
k_cu_seqlens=cu_seqlens,
max_seqlen=max_seqlen,
)
if self.use_tensor_parallel:
attn_out, _ = self.wo(attn_out)
else:
attn_out = self.wo(attn_out)
return attn_out
def forward(
self,
@@ -618,6 +448,7 @@ class MoonVitEncoderLayer(nn.Module):
cu_seqlens: torch.Tensor,
rope_freqs_cis: Union[torch.Tensor, None] = None,
max_seqlen: Optional[int] = None,
forward_metadata: Optional[VisionAttentionMetadata] = None,
) -> torch.Tensor:
"""
Args:
@@ -628,13 +459,14 @@ class MoonVitEncoderLayer(nn.Module):
"""
residual = hidden_states
hidden_states = self.norm0(hidden_states)
attn_out = self.attention_qkvpacked(
attn_out = self.attn(
hidden_states,
cu_seqlens,
rope_freqs_cis=rope_freqs_cis,
cu_seqlens=cu_seqlens,
position_embeddings=rope_freqs_cis,
forward_metadata=forward_metadata,
max_seqlen=max_seqlen,
)
hidden_states = residual + attn_out
hidden_states = residual + attn_out.view(residual.shape)
residual = hidden_states
hidden_states = self.mlp(self.norm1(hidden_states))
@@ -651,7 +483,6 @@ class MoonVitEncoder(nn.Module):
block_cfg: dict,
prefix: str = "",
use_data_parallel: bool = False,
use_tensor_parallel: bool = False,
) -> None:
super().__init__()
@@ -663,7 +494,6 @@ class MoonVitEncoder(nn.Module):
MoonVitEncoderLayer(
prefix=add_prefix(f"blocks.{layer_idx}", prefix),
use_data_parallel=use_data_parallel,
use_tensor_parallel=use_tensor_parallel,
**block_cfg,
)
for layer_idx in range(num_layers)
@@ -689,12 +519,17 @@ class MoonVitEncoder(nn.Module):
if max_seqlen is None:
max_seqlen = (grid_hw[:, 0] * grid_hw[:, 1]).max().item()
forward_metadata = prepare_vision_attention_metadata(
cu_seqlens, device=hidden_states.device
)
for _, block in enumerate(self.blocks):
hidden_states = block(
hidden_states,
cu_seqlens,
rope_freqs_cis=rope_freqs_cis,
max_seqlen=max_seqlen,
forward_metadata=forward_metadata,
)
hidden_states = self.final_layernorm(hidden_states)
@@ -769,7 +604,6 @@ class MoonVitPretrainedModel(PreTrainedModel):
config: MoonViTConfig,
prefix: str = "",
use_data_parallel: bool = False,
use_tensor_parallel: bool = False,
*inputs,
**kwargs,
):
@@ -795,11 +629,9 @@ class MoonVitPretrainedModel(PreTrainedModel):
"mlp_dim": config.intermediate_size,
"activation": GELUTanh(),
"attn_bias": True,
"attn_implementation": config._attn_implementation,
},
prefix=add_prefix("encoder", prefix),
use_data_parallel=use_data_parallel,
use_tensor_parallel=use_tensor_parallel,
)
def forward(
python/sglang/srt/models/nvila.py
+1 -3
View File
@@ -195,12 +195,10 @@ class NVILAForConditionalGeneration(nn.Module):
for x, block_size in zip(vision_features_list, block_sizes)
]
vision_features = torch.stack(
vision_features = torch.cat(
[einops.rearrange(x, "1 c h w -> (h w) c") for x in vision_features_list]
)
vision_features = einops.rearrange(vision_features, "n p d -> (n p) d")
return vision_features
def load_weights(self, weights: Iterable[tuple[str, Tensor]]) -> None: