Skip to content
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

misc: use the new plan/run API for unittests #467

Merged
merged 3 commits into from
Aug 26, 2024
Merged

misc: use the new plan/run API for unittests #467

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e707665
upd
yzh119 Aug 26, 2024
b639511
upd
yzh119 Aug 26, 2024
a11ee31
fix
yzh119 Aug 26, 2024
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
6 changes: 6 additions & 0 deletions docs/api/python/cascade.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,11 +28,17 @@ Cascade Attention Wrapper Classes
.. autoclass:: MultiLevelCascadeAttentionWrapper
:members:

.. automethod:: __init__

:exclude-members: begin_forward, end_forward, forward, forward_return_lse

.. autoclass:: BatchDecodeWithSharedPrefixPagedKVCacheWrapper
:members:

.. automethod:: __init__

.. autoclass:: BatchPrefillWithSharedPrefixPagedKVCacheWrapper
:members:

.. automethod:: __init__

2 changes: 2 additions & 0 deletions docs/api/python/decode.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,8 @@ Batch Decoding
:members:

.. automethod:: __init__

:exclude-members: begin_forward, end_forward, forward, forward_return_lse

.. autoclass:: CUDAGraphBatchDecodeWithPagedKVCacheWrapper
:members:
Expand Down
4 changes: 4 additions & 0 deletions docs/api/python/prefill.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,11 @@ Batch Prefill/Append Attention

.. automethod:: __init__

:exclude-members: begin_forward, end_forward, forward, forward_return_lse

.. autoclass:: BatchPrefillWithRaggedKVCacheWrapper
:members:

.. automethod:: __init__

:exclude-members: begin_forward, end_forward, forward, forward_return_lse
4 changes: 3 additions & 1 deletion docs/api/python/sparse.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,4 +10,6 @@ Kernels for block sparse flashattention.
.. autoclass:: BlockSparseAttentionWrapper
:members:

.. automethod:: __init__
.. automethod:: __init__

:exclude-members: begin_forward, end_forward, forward, forward_return_lse
2 changes: 1 addition & 1 deletion docs/tutorials/kv_layout.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -155,7 +155,7 @@ When using multi-level `cascade inference <https://flashinfer.ai/2024/02/02/casc
the query and output are stored in ragged tensors, and KV-Cache of all levels are stored
in a unified Paged KV-Cache. Each level has a unique ``qo_indptr`` array which is the prefix sum of the
accumulated number of tokens to append in the subtree, as well as ``kv_page_indptr``, ``kv_page_indices``, and
``kv_last_page_len`` which has same semantics as in :ref:`<page-layout>` section. The following figure
``kv_last_page_len`` which has same semantics as in :ref:`page-layout` section. The following figure
introduce how to construct these data structures for append attention operation for 8 requests where we
treat their KV-Cache as 3 levels for prefix reuse:

Expand Down
22 changes: 9 additions & 13 deletions python/flashinfer/prefill.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -478,15 +478,14 @@ class BatchPrefillWithPagedKVCacheWrapper:
... num_kv_heads,
... head_dim,
... page_size,
... causal=True,
... )
>>> outputs = []
>>> for i in range(num_layers):
... q = q_at_layer[i]
... kv_cache = kv_cache_at_layer[i]
... # compute batch prefill attention, reuse auxiliary data structures
... o = prefill_wrapper.run(
... q, kv_cache, causal=True
... )
... o = prefill_wrapper.run(q, kv_cache)
... outputs.append(o)
...
>>> outputs[0].shape
Expand All @@ -513,20 +512,18 @@ class BatchPrefillWithPagedKVCacheWrapper:
... num_kv_heads,
... head_dim,
... page_size,
... mask
... custom_mask=mask,
... )
>>> outputs_custom_mask = []
>>> for i in range(num_layers):
... q = q_at_layer[i]
... kv_cache = kv_cache_at_layer[i]
... # compute batch prefill attention, reuse auxiliary data structures
... o_custom = prefill_wrapper.run(
... q, kv_cache
... )
... o_custom = prefill_wrapper.run(q, kv_cache)
... assert torch.allclose(o_custom, outputs[i], rtol=1e-3, atol=1e-3)
...



Note
----
To accelerate computation, FlashInfer's batch prefill/append attention operators
Expand Down Expand Up @@ -1161,17 +1158,16 @@ class BatchPrefillWithRaggedKVCacheWrapper:
... kv_indptr,
... num_qo_heads,
... num_kv_heads,
... head_dim
... head_dim,
... causal=True,
... )
>>> outputs = []
>>> for i in range(num_layers):
... q = q_at_layer[i]
... k = k_at_layer[i]
... v = v_at_layer[i]
... # compute batch prefill attention, reuse auxiliary data structures
... o = prefill_wrapper.run(
... q, k, v, causal=True
... )
... o = prefill_wrapper.run(q, k, v)
... outputs.append(o)
...
>>> outputs[0].shape
Expand All @@ -1195,7 +1191,7 @@ class BatchPrefillWithRaggedKVCacheWrapper:
... num_qo_heads,
... num_kv_heads,
... head_dim,
... mask
... custom_mask=mask
... )
>>> outputs_custom_mask = []
>>> for i in range(num_layers):
Expand Down
53 changes: 16 additions & 37 deletions python/tests/test_batch_decode_kernels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,33 +65,23 @@ def test_batch_decode_with_paged_kv_cache(

workspace_buffer = torch.empty(32 * 1024 * 1024, dtype=torch.int8).to(0)
wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, kv_layout)
wrapper.begin_forward(
wrapper.plan(
kv_indptr,
kv_indices,
kv_last_page_len,
num_qo_heads,
num_kv_heads,
head_dim,
page_size,
"NONE",
logits_soft_cap=logits_soft_cap,
pos_encoding_mode=pos_encoding_mode,
data_type=kv_dtype,
q_data_type=q_dtype,
)
if return_lse:
o, _ = wrapper.forward_return_lse(
q,
kv_data.to(kv_dtype),
pos_encoding_mode=pos_encoding_mode,
logits_soft_cap=logits_soft_cap,
)
o, _ = wrapper.run_return_lse(q, kv_data.to(kv_dtype))
else:
o = wrapper.forward(
q,
kv_data.to(kv_dtype),
pos_encoding_mode=pos_encoding_mode,
logits_soft_cap=logits_soft_cap,
)
o = wrapper.run(q, kv_data.to(kv_dtype))

for i in range(batch_size):
perm_dims = [0, 2, 1, 3] if kv_layout == "HND" else [0, 1, 2, 3]
Expand Down Expand Up @@ -186,33 +176,23 @@ def test_batch_decode_with_tuple_paged_kv_cache(

workspace_buffer = torch.empty(32 * 1024 * 1024, dtype=torch.int8).to(0)
wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, kv_layout)
wrapper.begin_forward(
wrapper.plan(
kv_indptr,
kv_indices,
kv_last_page_len,
num_qo_heads,
num_kv_heads,
head_dim,
page_size,
"NONE",
logits_soft_cap=logits_soft_cap,
pos_encoding_mode=pos_encoding_mode,
data_type=kv_dtype,
q_data_type=q_dtype,
)
if return_lse:
o, _ = wrapper.forward_return_lse(
q,
tuple(map(lambda _: _.to(kv_dtype), kv_data)),
pos_encoding_mode=pos_encoding_mode,
logits_soft_cap=logits_soft_cap,
)
o, _ = wrapper.run_return_lse(q, tuple(map(lambda _: _.to(kv_dtype), kv_data)))
else:
o = wrapper.forward(
q,
tuple(map(lambda _: _.to(kv_dtype), kv_data)),
pos_encoding_mode=pos_encoding_mode,
logits_soft_cap=logits_soft_cap,
)
o = wrapper.run(q, tuple(map(lambda _: _.to(kv_dtype), kv_data)))

k_cache, v_cache = kv_data
for i in range(batch_size):
Expand Down Expand Up @@ -313,48 +293,47 @@ def test_cuda_graph_batch_decode_with_paged_kv_cache(
kv_last_page_device_buffer,
kv_layout,
)
wrapper.begin_forward(
wrapper.plan(
kv_indptr_host_warmup,
kv_indices_host_warmup,
kv_last_page_len_host_warmup,
num_qo_heads,
num_kv_heads,
head_dim,
page_size,
"NONE",
data_type=kv_dtype,
pos_encoding_mode=pos_encoding_mode,
q_data_type=q_dtype,
)
# warmup
s = torch.cuda.Stream()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
for _ in range(3):
o = wrapper.forward(q, kv_data_dtype, pos_encoding_mode=pos_encoding_mode)
o = wrapper.run(q, kv_data_dtype)
torch.cuda.current_stream().wait_stream(s)

# capture
g = torch.cuda.CUDAGraph()
with torch.cuda.graph(g):
o = wrapper.forward(q, kv_data_dtype, pos_encoding_mode=pos_encoding_mode)
wrapper.end_forward()
o = wrapper.run(q, kv_data_dtype)

# replay multiple times
for i in range(1, min(4, num_pages_per_seq)):
kv_indptr_host = torch.arange(0, batch_size + 1).int() * i
kv_indices_host = torch.arange(0, i * batch_size).int()
kv_last_page_len_host = torch.full((batch_size,), page_size, dtype=torch.int32)

wrapper.begin_forward(
wrapper.plan(
kv_indptr_host,
kv_indices_host,
kv_last_page_len_host,
num_qo_heads,
num_kv_heads,
head_dim,
page_size,
"NONE",
data_type=kv_dtype,
pos_encoding_mode=pos_encoding_mode,
q_data_type=q_dtype,
)
g.replay()
Expand All @@ -366,16 +345,16 @@ def test_cuda_graph_batch_decode_with_paged_kv_cache(
(batch_size,), (kv_len - 1) % page_size + 1, dtype=torch.int32
)

wrapper.begin_forward(
wrapper.plan(
kv_indptr_host,
kv_indices_host,
kv_last_page_len_host,
num_qo_heads,
num_kv_heads,
head_dim,
page_size,
"NONE",
data_type=kv_dtype,
pos_encoding_mode=pos_encoding_mode,
q_data_type=q_dtype,
)
g.replay()
Expand Down
Loading