Skip to content

Hybrid chunked cache update #500

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 2 commits into
base: release/v1.20.0
Choose a base branch
from
+85 −34
Draft

Hybrid chunked cache update #500

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8e0aa38
Cache update for Hybrid chunk attention
quic-amitraj Jul 8, 2025
b4ad73c
Ruff fix
quic-amitraj Jul 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
119 changes: 85 additions & 34 deletions QEfficient/transformers/cache_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -442,49 +442,100 @@ def update(
k_out, v_out = key_states, value_states

else:
layer_ctx_len = self.key_cache[layer_idx].shape[2]

position_ids = cache_kwargs.get("position_ids")
is_sliding_layer = torch.tensor(bool(self.is_sliding[layer_idx]))

# Update the position_ids to handle the sliding window
layer_ctx_len = self.key_cache[layer_idx].shape[2]
kv_position_ids = torch.where(
(~is_sliding_layer | (position_ids == -1)), position_ids, position_ids % (layer_ctx_len - 1)
)
scatter_position_id = position_ids
if is_sliding_layer:
scatter_position_id = position_ids % layer_ctx_len

kv_position_ids = torch.where(
is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1) * 2),
(position_ids + 1) % layer_ctx_len,
kv_position_ids,
)
self.key_cache[layer_idx] = CtxScatterFunc.apply(self.key_cache[layer_idx], scatter_position_id, key_states)

valid_mask = (kv_position_ids != -1).unsqueeze(1).unsqueeze(-1)
key_states = torch.where(valid_mask == 1, key_states, torch.zeros_like(key_states))
value_states = torch.where(valid_mask == 1, value_states, torch.zeros_like(value_states))
self.key_cache[layer_idx] = CtxScatterFunc.apply(self.key_cache[layer_idx], kv_position_ids, key_states)
self.value_cache[layer_idx] = CtxScatterFunc.apply(
self.value_cache[layer_idx], kv_position_ids, value_states
self.value_cache[layer_idx], scatter_position_id, value_states
)
k_out, v_out = self.key_cache[layer_idx], self.value_cache[layer_idx]
if is_sliding_layer:
kv_position_ids = torch.where(
(~is_sliding_layer | (position_ids == -1)), position_ids, position_ids % (layer_ctx_len)
)

kv_position_ids = torch.where(
is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1) * 2),
(position_ids + 1) % layer_ctx_len,
kv_position_ids,
)
ctx_len = min(layer_ctx_len, k_out.shape[2])
ctx_indices = torch.arange(ctx_len)[None, None, ...]
gather_limit = kv_position_ids.max(1, keepdim=True).values.unsqueeze(1)
invalid_mask = ctx_indices > gather_limit

if torch.onnx.is_in_onnx_export():
invalid_idx_value = torch.iinfo(torch.int32).max
else:
invalid_idx_value = 0

ctx_indices = torch.where(invalid_mask, invalid_idx_value, ctx_indices)

# Rolling indices for sliding window
all_indices = torch.arange(layer_ctx_len) + kv_position_ids.max() + 1
rolling_indices = torch.where(all_indices > layer_ctx_len - 1, all_indices % layer_ctx_len, all_indices)
final_indices = torch.where(
(is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), rolling_indices, ctx_indices
)

# print("k_out", k_out.shape, "v_out", v_out.shape)
k_out = torch.where(
self.key_cache[3].shape[-2] > 8192,
self.update_hybrid_cache_k_out(k_out, final_indices),
self.update_dynamic_cache_k_out(k_out, ctx_indices),
)
v_out = torch.where(
self.key_cache[3].shape[-2] > 8192,
self.update_hybrid_cache_v_out(
v_out, final_indices, invalid_mask, layer_ctx_len, position_ids, is_sliding_layer
),
self.update_dynamic_cache_v_out(v_out, ctx_indices, invalid_mask),
)

# Original Gather
ctx_len = min(layer_ctx_len, k_out.shape[2])
ctx_indices = torch.arange(ctx_len)[None, None, ...]
gather_limit = kv_position_ids.max(1, keepdim=True).values.unsqueeze(1)
invalid_mask = ctx_indices > gather_limit
if torch.onnx.is_in_onnx_export():
invalid_idx_value = torch.iinfo(torch.int32).max
else:
invalid_idx_value = 0
ctx_indices = torch.where(invalid_mask, invalid_idx_value, ctx_indices)
ctx_len = k_out.shape[2]
ctx_indices = torch.arange(ctx_len)[None, None, ...]
gather_limit = position_ids.max(1, keepdim=True).values.unsqueeze(1)
invalid_mask = ctx_indices > gather_limit

if torch.onnx.is_in_onnx_export():
invalid_idx_value = torch.iinfo(torch.int32).max
else:
invalid_idx_value = 0

ctx_indices = torch.where(invalid_mask, invalid_idx_value, ctx_indices)

k_out = self.update_dynamic_cache_k_out(k_out, ctx_indices)
v_out = self.update_dynamic_cache_v_out(v_out, ctx_indices, invalid_mask)

# Rolling indices for sliding window
all_indices = torch.arange(layer_ctx_len) + kv_position_ids.max() + 1
rolling_indices = torch.where(all_indices > layer_ctx_len - 1, all_indices % layer_ctx_len, all_indices)
final_indices = torch.where(
(is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), rolling_indices, ctx_indices
)
k_out = CtxGatherFunc.apply(k_out, final_indices)
v_out = CtxGatherFunc.apply(v_out, final_indices)
ctx_v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
v_out = torch.where((is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), v_out, ctx_v_out)
return k_out, v_out

def update_hybrid_cache_k_out(self, k_out, final_indices):
k_out = CtxGatherFunc.apply(k_out, final_indices)
return k_out

def update_hybrid_cache_v_out(
self, v_out, final_indices, invalid_mask, layer_ctx_len, position_ids, is_sliding_layer
):
v_out = CtxGatherFunc.apply(v_out, final_indices)
ctx_v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
v_out = torch.where((is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), v_out, ctx_v_out)

return v_out

def update_dynamic_cache_k_out(self, k_out, ctx_indices):
k_out = CtxGatherFunc.apply(k_out, ctx_indices)
return k_out

def update_dynamic_cache_v_out(self, v_out, ctx_indices, invalid_mask):
v_out = CtxGatherFunc.apply(v_out, ctx_indices)
v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
return v_out