Cccccz's picture
Add files using upload-large-folder tool
d2b26ce verified
Raw
History Blame Contribute Delete
14.1 kB
# Copyright (c) 2025 SandAI. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Utility functions for cache management.
"""
import math
import torch
from typing import Dict, List, Tuple, Optional, Any
from inference.common import PackedCrossAttnParams
def generate_dynamic_kv_range(
tracker,
current_chunk_id: int,
x_chunks_keys: List[int],
chunk_token_nums: int,
near_clean_chunk_idx: int = -1
) -> torch.Tensor:
"""
Generate dynamic KV ranges for chunks after compression.
This function computes the KV range each chunk should attend to,
taking into account the compressed KV cache layout.
Args:
tracker: ChunkKVRangeTracker instance managing chunk ranges
current_chunk_id: The chunk being processed
x_chunks_keys: List of all chunk keys being processed
chunk_token_nums: Number of tokens per chunk
near_clean_chunk_idx: Index of the nearly-clean chunk (-1 if not present)
Returns:
Tensor of shape [num_chunks, 2] with KV ranges for each chunk
"""
kv_ranges = []
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Process normal chunks (excluding near_clean_chunk)
normal_chunks = [chunk_id for chunk_id in x_chunks_keys if chunk_id != near_clean_chunk_idx]
for chunk_id in normal_chunks:
# Normal chunk: needs to see itself and all previous chunks
all_chunk_ids = tracker.get_all_chunk_ids() + list(normal_chunks)
chunks_to_include = [cid for cid in all_chunk_ids if cid <= chunk_id]
# Calculate based on actual compressed ranges in tracker
total_tokens = 0
for cid in chunks_to_include:
if cid in tracker.get_all_chunk_ids():
# Use compressed actual range
s, e = tracker.get_range(cid)
total_tokens = max(total_tokens, e)
else:
# Newly entered chunk not yet registered, but size is known
total_tokens += chunk_token_nums
range_start = 0
range_end = total_tokens
kv_ranges.append([range_start, range_end])
# Handle near_clean_chunk (always last if present)
if near_clean_chunk_idx != -1:
# Calculate end position of last normal chunk
last_normal_chunk_end = 0
all_chunk_ids = tracker.get_all_chunk_ids() + normal_chunks
for cid in all_chunk_ids:
if cid in tracker.get_all_chunk_ids():
s, e = tracker.get_range(cid)
last_normal_chunk_end = max(last_normal_chunk_end, e)
else:
# Newly entered chunk not yet registered
last_normal_chunk_end += chunk_token_nums
# near_clean_chunk range: (last_normal_chunk_end, last_normal_chunk_end + chunk_token_nums]
range_start = last_normal_chunk_end
range_end = last_normal_chunk_end + chunk_token_nums
kv_ranges.append([range_start, range_end])
return torch.tensor(kv_ranges, device=device, dtype=torch.int32)
def identify_compressible_chunks(
tracker,
chunk_start: int,
transport_input,
chunk_denoise_count: Dict[int, int],
chunk_offset: int = 0
) -> Tuple[List[int], List[int]]:
"""
Identify which chunks can be compressed and which should remain active.
A chunk can be compressed if:
- It's a prefix video chunk (always clean)
- It's a generated chunk that has completed all denoising steps
Args:
tracker: ChunkKVRangeTracker instance
chunk_start: Current chunk being processed
transport_input: Transport input containing chunk info
chunk_denoise_count: Dictionary mapping chunk_id to denoising steps completed
chunk_offset: Number of prefix video chunks
Returns:
Tuple of (clean_chunk_ids, active_chunk_ids)
"""
all_chunk_ids = tracker.get_all_chunk_ids()
clean_chunks = []
for cid in all_chunk_ids:
if cid < chunk_offset:
# Prefix video chunks are always clean
clean_chunks.append(cid)
elif cid <= chunk_start:
# Generated chunks need to check denoising completion
if chunk_denoise_count[cid] == transport_input.num_steps:
clean_chunks.append(cid)
active_chunks = [cid for cid in all_chunk_ids if cid not in clean_chunks]
return clean_chunks, active_chunks
def check_compress_condition(
tracker,
total_cache_len: int,
chunk_num: int,
chunk_start: int,
transport_input,
chunk_denoise_count: Dict[int, int],
window_size: int = 4
) -> bool:
"""
Check if KV cache compression should be triggered.
Compression is triggered when:
1. Cache is full (next_free_idx >= total_cache_len)
2. More chunks are yet to enter (registered_count < chunk_num)
3. Next chunk is about to enter (last chunk's steps == num_steps/window_size)
Args:
tracker: ChunkKVRangeTracker instance
total_cache_len: Total cache capacity in tokens
chunk_num: Total number of chunks
chunk_start: Current chunk being processed
transport_input: Transport input containing parameters
chunk_denoise_count: Dictionary mapping chunk_id to denoising steps
window_size: Window size for denoising stages (default: 4)
Returns:
True if compression should be performed, False otherwise
"""
all_chunk_ids = tracker.get_all_chunk_ids()
if len(all_chunk_ids) == 0:
return False
registered_chunk_count = len(all_chunk_ids)
cache_full = tracker.next_free_idx >= total_cache_len
has_more_chunks = registered_chunk_count < chunk_num
last_chunk_id = all_chunk_ids[-1]
# Calculate steps per stage
steps_per_stage = transport_input.num_steps // window_size
next_chunk_will_enter = chunk_denoise_count[last_chunk_id] == steps_per_stage
should_compress = cache_full and has_more_chunks and next_chunk_will_enter
return should_compress
def get_embedding_and_meta_with_chunk_info(
model_self,
x: torch.Tensor,
t: torch.Tensor,
y: torch.Tensor,
caption_dropout_mask,
xattn_mask,
kv_range: torch.Tensor,
**kwargs
) -> tuple:
"""
Compute embeddings and meta information with chunk-aware processing.
This is a unified version of the get_embedding_and_meta function that
properly handles chunk-based processing with dynamic KV ranges.
Args:
model_self: The DiT model instance
x: Input tensor [N, C, T, H, W]
t: Timestep tensor [N, range_num]
y: Text conditioning tensor
caption_dropout_mask: Dropout mask for captions
xattn_mask: Cross-attention mask
kv_range: KV range tensor
**kwargs: Additional arguments including:
- range_num: Total number of chunks
- denoising_range_num: Number of chunks being denoised
- slice_point: Starting chunk index
- start_chunk_id: First chunk to process
- end_chunk_id: Last chunk to process (exclusive)
- distill_nearly_clean_chunk: Whether to add nearly-clean chunk
- chunk_token_nums: Tokens per chunk
- chunk_width: Width of each chunk in frames
- num_steps: Total denoising steps
Returns:
Tuple of (x, condition, condition_map, rope, y_xattn_flat, xattn_mask_cuda,
H, W, ardf_meta, cross_attn_params)
"""
# ========== Part 1: Embed x ==========
x = model_self.x_embedder(x) # [N, C, T, H, W]
batch_size, _, T, H, W = x.shape
# Prepare necessary variables
range_num = kwargs["range_num"]
denoising_range_num = kwargs["denoising_range_num"]
slice_point = kwargs.get("slice_point", 0)
frame_in_range = T // denoising_range_num
# distill_nearly_clean_chunk adds one extra chunk
T_total = (range_num + kwargs.get("distill_nearly_clean_chunk", False)) * frame_in_range
# ========== Part 2: Compute rotary positional embedding ==========
rescale_factor = math.sqrt((H * W) / (16 * 16))
rope = model_self.rope.get_embed(
shape=[T_total, H, W],
ref_feat_shape=[T_total, H / rescale_factor, W / rescale_factor]
)
# Rope shape: (T*H*W, head_dim) - cut to current chunk range
rope = rope[
kwargs["start_chunk_id"] * frame_in_range * H * W :
kwargs["end_chunk_id"] * frame_in_range * H * W
]
# ========== Part 3: Embed t ==========
assert t.shape[0] == batch_size, f"Invalid t shape: {t.shape[0]} != {batch_size}"
assert t.shape[1] == denoising_range_num, f"Invalid t shape: {t.shape[1]} != {denoising_range_num}"
t_flat = t.flatten() # (N * denoising_range_num,)
t = model_self.t_embedder(t_flat) # (N, D)
if model_self.engine_config.distill:
distill_dt_scalar = 2
if kwargs["num_steps"] == 12:
base_chunk_step = 4
distill_dt_factor = base_chunk_step / kwargs["distill_interval"] * distill_dt_scalar
else:
distill_dt_factor = kwargs["num_steps"] / 4 * distill_dt_scalar
distill_dt = torch.ones_like(t_flat) * distill_dt_factor
distill_dt_embed = model_self.t_embedder(distill_dt)
t = t + distill_dt_embed
t = t.reshape(batch_size, denoising_range_num, -1) # (N, range_num, D)
# ========== Part 4: Embed y, prepare condition and y_xattn_flat ==========
y_xattn, y_adaln = model_self.y_embedder(y, model_self.training, caption_dropout_mask)
assert xattn_mask is not None
xattn_mask = xattn_mask.squeeze(1).squeeze(1)
# condition: (N, range_num, D)
y_adaln = y_adaln.squeeze(1) # (N, D)
condition = t + y_adaln.unsqueeze(1)
assert condition.shape[0] == batch_size
assert condition.shape[1] == denoising_range_num
seqlen_per_chunk = (T * H * W) // denoising_range_num
condition_map = torch.arange(batch_size * denoising_range_num, device=x.device)
condition_map = torch.repeat_interleave(condition_map, seqlen_per_chunk)
condition_map = condition_map.reshape(batch_size, -1).transpose(0, 1).contiguous()
# y_xattn_flat: (total_token, D)
y_xattn_flat = torch.masked_select(
y_xattn.squeeze(1),
xattn_mask.unsqueeze(-1).bool()
).reshape(-1, y_xattn.shape[-1])
xattn_mask_for_cuda_graph = None
# ========== Part 5: Prepare cross_attn_params ==========
xattn_mask = xattn_mask.reshape(xattn_mask.shape[0], -1)
y_index = torch.sum(xattn_mask, dim=-1)
clip_token_nums = H * W * frame_in_range
cu_seqlens_q = torch.Tensor(
[0] + ([clip_token_nums] * denoising_range_num * batch_size)
).to(torch.int64).to(x.device)
cu_seqlens_k = torch.cat(
[y_index.new_tensor([0]), y_index]
).to(torch.int64).to(x.device)
cu_seqlens_q = cu_seqlens_q.cumsum(-1).to(torch.int32)
cu_seqlens_k = cu_seqlens_k.cumsum(-1).to(torch.int32)
assert cu_seqlens_q.shape == cu_seqlens_k.shape, \
f"cu_seqlens_q.shape: {cu_seqlens_q.shape}, cu_seqlens_k.shape: {cu_seqlens_k.shape}"
xattn_q_ranges = torch.cat(
[cu_seqlens_q[:-1].unsqueeze(1), cu_seqlens_q[1:].unsqueeze(1)], dim=1
)
xattn_k_ranges = torch.cat(
[cu_seqlens_k[:-1].unsqueeze(1), cu_seqlens_k[1:].unsqueeze(1)], dim=1
)
assert xattn_q_ranges.shape == xattn_k_ranges.shape, \
f"xattn_q_ranges.shape: {xattn_q_ranges.shape}, xattn_k_ranges.shape: {xattn_k_ranges.shape}"
cross_attn_params = PackedCrossAttnParams(
q_ranges=xattn_q_ranges,
kv_ranges=xattn_k_ranges,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_kv=cu_seqlens_k,
max_seqlen_q=clip_token_nums,
max_seqlen_kv=model_self.caption_max_length,
)
# ========== Part 6: Prepare core_attn related q/kv range ==========
q_range = torch.cat(
[cu_seqlens_q[:-1].unsqueeze(1), cu_seqlens_q[1:].unsqueeze(1)], dim=1
)
flat_kv = torch.unique(kv_range, sorted=True)
max_seqlen_k = (flat_kv[-1] - flat_kv[0]).cpu().item()
ardf_meta = dict(
clip_token_nums=clip_token_nums,
slice_point=slice_point,
range_num=range_num,
denoising_range_num=denoising_range_num,
q_range=q_range,
k_range=kv_range,
max_seqlen_q=clip_token_nums,
max_seqlen_k=max_seqlen_k,
)
return (x, condition, condition_map, rope, y_xattn_flat,
xattn_mask_for_cuda_graph, H, W, ardf_meta, cross_attn_params)
def compute_chunk_token_nums(
transport_input,
model_config,
chunk_width: int
) -> int:
"""
Calculate the number of tokens in one chunk.
Args:
transport_input: Transport input containing latent dimensions
model_config: Model configuration
chunk_width: Number of frames per chunk
Returns:
Number of tokens per chunk
"""
patch_size = model_config.patch_size
latent_h = transport_input.latent_size[3] // patch_size
latent_w = transport_input.latent_size[4] // patch_size
return chunk_width * latent_h * latent_w
def get_latent_spatial_dims(
transport_input,
model_config
) -> Tuple[int, int]:
"""
Get the spatial dimensions of latent in patch units.
Args:
transport_input: Transport input containing latent dimensions
model_config: Model configuration
Returns:
Tuple of (height_patches, width_patches)
"""
patch_size = model_config.patch_size
h = transport_input.latent_size[3] // patch_size
w = transport_input.latent_size[4] // patch_size
return h, w