[Bugfix][CI] fix typos (#34934)
Signed-off-by: 1195343015 <1195343015@qq.com> Signed-off-by: Jiayi Yan <66017932+1195343015@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
co-authored by
Harry Mellor
parent
8c760b6ab6
commit
6a895197fa
@@ -2540,7 +2540,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
|
||||
)
|
||||
# workspace
|
||||
# |------- N tokens --------|--------- N*dcp_size tokens ----------|
|
||||
# |<- use for loca_gather ->|<--------- use for allgather -------->|
|
||||
# |<- use for local_gather ->|<--------- use for allgather -------->|
|
||||
allgather_offset = workspace.shape[0] // (dcp_world_size + 1)
|
||||
assert allgather_offset * (dcp_world_size + 1) == workspace.shape[0]
|
||||
assert toks <= allgather_offset
|
||||
|
||||
@@ -394,5 +394,5 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):
|
||||
|
||||
def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None:
|
||||
# No support for LoRA in flashinfer_cutlass_fused_moe.
|
||||
# See TODOs in flashinfer functions runMoe and runMoeMinLantency.
|
||||
# See TODOs in flashinfer functions runMoe and runMoeMinLatency.
|
||||
raise NotImplementedError("LoRA is not supported for flashinfer_cutlass_moe")
|
||||
|
||||
@@ -409,7 +409,7 @@ def batched_fused_marlin_moe(
|
||||
Note that the moe_align_block_size function indicates,
|
||||
- What rows of the A matrix (hidden_states) to access during the
|
||||
matmul, via sorted_ids output.
|
||||
- What expert_id to use for each block matmul, via expert_ids ouptut.
|
||||
- What expert_id to use for each block matmul, via expert_ids output.
|
||||
|
||||
In the batched version, the tokens are already grouped/batched by experts
|
||||
they subscribe to. Due to this, we can represent the batched hidden_states
|
||||
|
||||
@@ -606,7 +606,7 @@ class FusedMoEExperts(ABC):
|
||||
"""
|
||||
Whether the kernel supports deployment in particular parallel config.
|
||||
|
||||
Can be overriden if a kernel does not support EP, SP or some other
|
||||
Can be overridden if a kernel does not support EP, SP or some other
|
||||
configuration.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
@@ -620,7 +620,7 @@ class FusedMoEExperts(ABC):
|
||||
"""
|
||||
Whether the kernel supports a routing method (e.g. GroupedTopK).
|
||||
|
||||
Can be overriden by monolithic kernels that execute the router
|
||||
Can be overridden by monolithic kernels that execute the router
|
||||
in addition to the experts if certain routers are not supported.
|
||||
"""
|
||||
return True
|
||||
@@ -633,7 +633,7 @@ class FusedMoEExperts(ABC):
|
||||
"""
|
||||
Whether a kernel supports a particular dtype for router logits input.
|
||||
|
||||
Can be overriden by monolithic kernels that execute the router
|
||||
Can be overridden by monolithic kernels that execute the router
|
||||
in addition to the experts if certain dtypes are not supported.
|
||||
"""
|
||||
return True
|
||||
|
||||
@@ -1502,10 +1502,10 @@ class RowParallelLinear(LinearBase):
|
||||
if self.input_is_parallel:
|
||||
input_parallel = input_
|
||||
else:
|
||||
splitted_input = split_tensor_along_last_dim(
|
||||
split_input = split_tensor_along_last_dim(
|
||||
input_, num_partitions=self.tp_size
|
||||
)
|
||||
input_parallel = splitted_input[self.tp_rank].contiguous()
|
||||
input_parallel = split_input[self.tp_rank].contiguous()
|
||||
|
||||
# Matrix multiply.
|
||||
assert self.quant_method is not None
|
||||
|
||||
@@ -35,7 +35,7 @@ class MultiHeadLatentAttentionWrapper(PluggableLayer):
|
||||
"""Pluggable MLA layer which allows OOT backends to add
|
||||
custom implementations of the outer MLA layer (including rope & o_proj).
|
||||
Note that currently oot platforms can still use CustomOp.register_oot to
|
||||
replace MLA layer entirly, although we use PluggableLayer to register
|
||||
replace MLA layer entirely, although we use PluggableLayer to register
|
||||
this layer now.
|
||||
|
||||
This class takes positions and hidden_states as input.
|
||||
|
||||
@@ -191,7 +191,7 @@ class CompressedTensorsConfig(QuantizationConfig):
|
||||
"""
|
||||
Helper function to update target_scheme_map
|
||||
since linear layers get fused into FusedMoE
|
||||
targetting 'Linear' needs to also match
|
||||
targeting 'Linear' needs to also match
|
||||
FusedMoE modules.
|
||||
"""
|
||||
if (
|
||||
|
||||
+1
-1
@@ -2445,7 +2445,7 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
w2_scale=layer.w2_weight_scale, # group scale
|
||||
g1_alphas=layer.w13_weight_chan_scale,
|
||||
g2_alphas=layer.w2_weight_chan_scale,
|
||||
per_act_token_quant=True, # always use dynamc per-token
|
||||
per_act_token_quant=True, # always use dynamic per-token
|
||||
per_out_ch_quant=True, # always use per-channel
|
||||
)
|
||||
|
||||
|
||||
@@ -261,7 +261,7 @@ class CPUAWQLinearMethod(LinearMethodBase):
|
||||
|
||||
zeros = pack_cols(zeros, bits, group_num, output_size).contiguous()
|
||||
# make 16 output channel as a block and transpose to
|
||||
# the make the block contigous
|
||||
# the make the block contiguous
|
||||
weight = pack_cols(weight, bits, input_size, output_size)
|
||||
weight = (
|
||||
weight.view(input_size, -1, 16 // pack_factor)
|
||||
|
||||
@@ -199,7 +199,7 @@ class TorchAOConfig(QuantizationConfig):
|
||||
|
||||
@classmethod
|
||||
def from_config_dict_json(cls, config_dict_json: str) -> "TorchAOConfig":
|
||||
"""Iniitalize class from a config_dict json string, got from
|
||||
"""Initialize class from a config_dict json string, got from
|
||||
torchao_config_object = some AOBaseConfig object
|
||||
json.dumps(config_to_dict(torchao_config_object))
|
||||
"""
|
||||
|
||||
@@ -255,7 +255,7 @@ def _flashinfer_fp8_blockscale_gemm_impl(
|
||||
|
||||
This batch-size-dependent selection is essential for maintaining model accuracy.
|
||||
Benchmarks on GSM8K show a significant accuracy gap (88% vs 95%) for DeepSeek-V3.1
|
||||
when using FlashInfer's DeepGEMM on M>=32. The M < 32 strategy fixes the accurracy
|
||||
when using FlashInfer's DeepGEMM on M>=32. The M < 32 strategy fixes the accuracy
|
||||
drop.
|
||||
|
||||
Args:
|
||||
|
||||
@@ -39,7 +39,7 @@ def query_machete_supported_group_sizes(act_type: torch.dtype) -> list[int]:
|
||||
|
||||
|
||||
def check_machete_supports_shape(
|
||||
in_features: int, out_featrues: int
|
||||
in_features: int, out_features: int
|
||||
) -> tuple[bool, str | None]:
|
||||
if in_features % MACHETE_PREPACKED_BLOCK_SHAPE[0] != 0:
|
||||
return (
|
||||
@@ -47,7 +47,7 @@ def check_machete_supports_shape(
|
||||
"Input features size must be divisible by "
|
||||
f"{MACHETE_PREPACKED_BLOCK_SHAPE[0]}",
|
||||
)
|
||||
if out_featrues % MACHETE_PREPACKED_BLOCK_SHAPE[1] != 0:
|
||||
if out_features % MACHETE_PREPACKED_BLOCK_SHAPE[1] != 0:
|
||||
return (
|
||||
False,
|
||||
"Output features size must be divisible by "
|
||||
|
||||
@@ -237,7 +237,7 @@ class ApplyRotaryEmb(CustomOp):
|
||||
Arguments of apply_rotary_emb() in vllm_flash_attn:
|
||||
x: [batch_size, seq_len, nheads, headdim]
|
||||
cos, sin: [seqlen_rotary, rotary_dim / 2]
|
||||
interleaved: defalut as False (Neox-style).
|
||||
interleaved: default as False (Neox-style).
|
||||
...
|
||||
"""
|
||||
interleaved = not self.is_neox_style
|
||||
@@ -259,7 +259,7 @@ class ApplyRotaryEmb(CustomOp):
|
||||
Arguments of apply_rotary() in flash_attn:
|
||||
x: [batch_size, seq_len, nheads, headdim]
|
||||
cos, sin: [seqlen_rotary, rotary_dim / 2]
|
||||
interleaved: defalut as False (Neox-style).
|
||||
interleaved: default as False (Neox-style).
|
||||
...
|
||||
"""
|
||||
interleaved = not self.is_neox_style
|
||||
|
||||
Reference in New Issue
Block a user