Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
norm
vllm
Commits
73b1705e
Commit
73b1705e
authored
Mar 25, 2024
by
zhuwenwen
Browse files
add fp8
parent
dd823f7f
Changes
3
Hide whitespace changes
Inline
Side-by-side
Showing
3 changed files
with
32 additions
and
32 deletions
+32
-32
csrc/attention/attention_kernels.cu
csrc/attention/attention_kernels.cu
+20
-20
csrc/cache_kernels.cu
csrc/cache_kernels.cu
+10
-10
csrc/dispatch_utils.h
csrc/dispatch_utils.h
+2
-2
No files found.
csrc/attention/attention_kernels.cu
View file @
73b1705e
...
@@ -739,16 +739,16 @@ void paged_attention_v1(
...
@@ -739,16 +739,16 @@ void paged_attention_v1(
}
else
{
}
else
{
TORCH_CHECK
(
false
,
"Unsupported data type: "
,
query
.
dtype
());
TORCH_CHECK
(
false
,
"Unsupported data type: "
,
query
.
dtype
());
}
}
//
} else if (kv_cache_dtype == "fp8_e5m2") {
}
else
if
(
kv_cache_dtype
==
"fp8_e5m2"
)
{
//
if (query.dtype() == at::ScalarType::Float) {
if
(
query
.
dtype
()
==
at
::
ScalarType
::
Float
)
{
//
CALL_V1_LAUNCHER_BLOCK_SIZE(float, uint8_t, true);
CALL_V1_LAUNCHER_BLOCK_SIZE
(
float
,
uint8_t
,
true
);
//
} else if (query.dtype() == at::ScalarType::Half) {
}
else
if
(
query
.
dtype
()
==
at
::
ScalarType
::
Half
)
{
//
CALL_V1_LAUNCHER_BLOCK_SIZE(uint16_t, uint8_t, true);
CALL_V1_LAUNCHER_BLOCK_SIZE
(
uint16_t
,
uint8_t
,
true
);
//
} else if (query.dtype() == at::ScalarType::BFloat16) {
}
else
if
(
query
.
dtype
()
==
at
::
ScalarType
::
BFloat16
)
{
//
CALL_V1_LAUNCHER_BLOCK_SIZE(__nv_bfloat16, uint8_t, true);
CALL_V1_LAUNCHER_BLOCK_SIZE
(
__nv_bfloat16
,
uint8_t
,
true
);
//
} else {
}
else
{
//
TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
TORCH_CHECK
(
false
,
"Unsupported data type: "
,
query
.
dtype
());
//
}
}
}
else
{
}
else
{
TORCH_CHECK
(
false
,
"Unsupported data type of kv cache: "
,
kv_cache_dtype
);
TORCH_CHECK
(
false
,
"Unsupported data type of kv cache: "
,
kv_cache_dtype
);
}
}
...
@@ -932,16 +932,16 @@ void paged_attention_v2(
...
@@ -932,16 +932,16 @@ void paged_attention_v2(
}
else
{
}
else
{
TORCH_CHECK
(
false
,
"Unsupported data type: "
,
query
.
dtype
());
TORCH_CHECK
(
false
,
"Unsupported data type: "
,
query
.
dtype
());
}
}
//
} else if (kv_cache_dtype == "fp8_e5m2") {
}
else
if
(
kv_cache_dtype
==
"fp8_e5m2"
)
{
//
if (query.dtype() == at::ScalarType::Float) {
if
(
query
.
dtype
()
==
at
::
ScalarType
::
Float
)
{
//
CALL_V2_LAUNCHER_BLOCK_SIZE(float, uint8_t, true);
CALL_V2_LAUNCHER_BLOCK_SIZE
(
float
,
uint8_t
,
true
);
//
} else if (query.dtype() == at::ScalarType::Half) {
}
else
if
(
query
.
dtype
()
==
at
::
ScalarType
::
Half
)
{
//
CALL_V2_LAUNCHER_BLOCK_SIZE(uint16_t, uint8_t, true);
CALL_V2_LAUNCHER_BLOCK_SIZE
(
uint16_t
,
uint8_t
,
true
);
//
} else if (query.dtype() == at::ScalarType::BFloat16) {
}
else
if
(
query
.
dtype
()
==
at
::
ScalarType
::
BFloat16
)
{
//
CALL_V2_LAUNCHER_BLOCK_SIZE(__nv_bfloat16, uint8_t, true);
CALL_V2_LAUNCHER_BLOCK_SIZE
(
__nv_bfloat16
,
uint8_t
,
true
);
//
} else {
}
else
{
//
TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
TORCH_CHECK
(
false
,
"Unsupported data type: "
,
query
.
dtype
());
//
}
}
}
else
{
}
else
{
TORCH_CHECK
(
false
,
"Unsupported data type of kv cache: "
,
kv_cache_dtype
);
TORCH_CHECK
(
false
,
"Unsupported data type of kv cache: "
,
kv_cache_dtype
);
}
}
...
...
csrc/cache_kernels.cu
View file @
73b1705e
...
@@ -254,14 +254,14 @@ void reshape_and_cache(
...
@@ -254,14 +254,14 @@ void reshape_and_cache(
}
else
if
(
key
.
dtype
()
==
at
::
ScalarType
::
BFloat16
)
{
}
else
if
(
key
.
dtype
()
==
at
::
ScalarType
::
BFloat16
)
{
CALL_RESHAPE_AND_CACHE
(
__nv_bfloat16
,
__nv_bfloat16
,
false
);
CALL_RESHAPE_AND_CACHE
(
__nv_bfloat16
,
__nv_bfloat16
,
false
);
}
}
//
} else if (kv_cache_dtype == "fp8_e5m2") {
}
else
if
(
kv_cache_dtype
==
"fp8_e5m2"
)
{
//
if (key.dtype() == at::ScalarType::Float) {
if
(
key
.
dtype
()
==
at
::
ScalarType
::
Float
)
{
//
CALL_RESHAPE_AND_CACHE(float, uint8_t, true);
CALL_RESHAPE_AND_CACHE
(
float
,
uint8_t
,
true
);
//
} else if (key.dtype() == at::ScalarType::Half) {
}
else
if
(
key
.
dtype
()
==
at
::
ScalarType
::
Half
)
{
//
CALL_RESHAPE_AND_CACHE(uint16_t, uint8_t, true);
CALL_RESHAPE_AND_CACHE
(
uint16_t
,
uint8_t
,
true
);
//
} else if (key.dtype() == at::ScalarType::BFloat16) {
}
else
if
(
key
.
dtype
()
==
at
::
ScalarType
::
BFloat16
)
{
//
CALL_RESHAPE_AND_CACHE(__nv_bfloat16, uint8_t, true);
CALL_RESHAPE_AND_CACHE
(
__nv_bfloat16
,
uint8_t
,
true
);
//
}
}
}
else
{
}
else
{
TORCH_CHECK
(
false
,
"Unsupported data type of kv cache: "
,
kv_cache_dtype
);
TORCH_CHECK
(
false
,
"Unsupported data type of kv cache: "
,
kv_cache_dtype
);
}
}
...
@@ -314,7 +314,7 @@ void convert_fp8_e5m2(
...
@@ -314,7 +314,7 @@ void convert_fp8_e5m2(
CALL_CONVERT_FP8_E5M2
(
float
,
uint8_t
);
CALL_CONVERT_FP8_E5M2
(
float
,
uint8_t
);
}
else
if
(
dst_cache
.
dtype
()
==
at
::
ScalarType
::
Half
)
{
}
else
if
(
dst_cache
.
dtype
()
==
at
::
ScalarType
::
Half
)
{
CALL_CONVERT_FP8_E5M2
(
uint16_t
,
uint8_t
);
CALL_CONVERT_FP8_E5M2
(
uint16_t
,
uint8_t
);
//
} else if (dst_cache.dtype() == at::ScalarType::BFloat16) {
}
else
if
(
dst_cache
.
dtype
()
==
at
::
ScalarType
::
BFloat16
)
{
//
CALL_CONVERT_FP8_E5M2(__nv_bfloat16, uint8_t);
CALL_CONVERT_FP8_E5M2
(
__nv_bfloat16
,
uint8_t
);
}
}
}
}
csrc/dispatch_utils.h
View file @
73b1705e
...
@@ -18,8 +18,8 @@
...
@@ -18,8 +18,8 @@
#define VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(...) \
#define VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(...) \
AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \
AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \
AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
\
//
AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)
AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)
#define VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(TYPE, NAME, ...) \
#define VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(TYPE, NAME, ...) \
AT_DISPATCH_SWITCH( \
AT_DISPATCH_SWITCH( \
...
...
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment