Skip to content

[Executorch][llm] Fix ring kv cache when used with quantized kv cache and sdpa #12143

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

Merged
merged 4 commits into from
Jul 3, 2025
Merged

[Executorch][llm] Fix ring kv cache when used with quantized kv cache and sdpa #12143

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
7b9ab92
[Executorch][llm] Make runner return error if execution was not succe…
kimishpatel Jul 1, 2025
09d57aa
[Executorch][llm] Make mask tensor float only for sdpa
kimishpatel Jul 1, 2025
e7a0c58
[Executorch][llm] Fix ring kv cache when used with quantized kv cache…
kimishpatel Jul 1, 2025
2530b33
Merge branch 'main' into gh/kimishpatel/196/orig
kimishpatel Jul 3, 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
14 changes: 13 additions & 1 deletion examples/models/llama/source_transformation/custom_kv_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,7 @@ def __init__(
head_dim,
cache_type: QuantizedCacheType = QuantizedCacheType.AffineSymmetric,
use_custom_update_cache_op: bool = False,
return_float_values: bool = True,
):
super().__init__()
if cache_type not in (
Expand All @@ -57,7 +58,7 @@ def __init__(
self.use_custom_update_cache_op = use_custom_update_cache_op
self.quantized_cache_dtype = torch.int8
self.cache_fp_type = torch.float32
self.return_float_values = True
self.return_float_values = return_float_values
self.max_context_length = max_context_length
cache_shape = (max_batch_size, max_context_length, n_heads, head_dim)
scale_shape = (max_batch_size, max_context_length, n_heads, 1)
Expand Down Expand Up @@ -400,6 +401,7 @@ def __init__(
head_dim,
cache_type: QuantizedCacheType = QuantizedCacheType.AffineSymmetric,
use_custom_update_cache_op: bool = False,
return_float_values: bool = True,
):
# Look at attention.py for explanation on why max_context_length * 2
super().__init__(
Expand All @@ -409,6 +411,7 @@ def __init__(
head_dim,
cache_type,
use_custom_update_cache_op,
return_float_values,
)
self.cache_positions_manager = CachePositionsManager(self.max_context_length)
self.is_ring_buffer = True
Expand Down Expand Up @@ -459,6 +462,7 @@ def from_quantized_kv_cache(
head_dim,
kv_cache.cache_type,
kv_cache.use_custom_update_cache_op,
kv_cache.return_float_values,
)


Expand Down Expand Up @@ -583,4 +587,12 @@ def replace_kv_cache_with_ring_kv_cache(module, layer_sizes):
# it is not doing causal attention
if "SDPACustom" in attention.SDPA.__class__.__name__:
attention.SDPA.use_attention_mask = True
# QuantizedSDPA has to store kv_cache in order to obtrain
# scales and zero points for k and v cache.
# So if we replcaed attention module's quantized kv cache with
# QuantizedRingKVCache then we also have to replace attention's
# SDPA module kv_cache so that it refers to the same kv_cache
if "QuantizedSDPA" in attention.SDPA.__class__.__name__:
attention.SDPA.use_attention_mask = True
attention.SDPA.kv_cache = attention.kv_cache
return module
Loading