chore(vllm): codespell for spell checking (#2820)
This commit is contained in:
@@ -178,7 +178,7 @@ class BlockSpaceManager:
|
||||
if len(block_table) < len(logical_blocks):
|
||||
if (self.block_sliding_window
|
||||
and len(block_table) >= self.block_sliding_window):
|
||||
# re-use a block
|
||||
# reuse a block
|
||||
block_table.append(block_table[len(block_table) %
|
||||
self.block_sliding_window])
|
||||
else:
|
||||
|
||||
@@ -158,7 +158,7 @@ class Scheduler:
|
||||
return len(self.waiting) + len(self.running) + len(self.swapped)
|
||||
|
||||
def _schedule(self) -> SchedulerOutputs:
|
||||
# Blocks that need to be swaped or copied before model execution.
|
||||
# Blocks that need to be swapped or copied before model execution.
|
||||
blocks_to_swap_in: Dict[int, int] = {}
|
||||
blocks_to_swap_out: Dict[int, int] = {}
|
||||
blocks_to_copy: Dict[int, List[int]] = {}
|
||||
|
||||
+1
-1
@@ -87,7 +87,7 @@ def add_lora(y: torch.Tensor,
|
||||
r = wb_t_all.size(-1)
|
||||
if buffer is None:
|
||||
# We set the buffer to be float32 by default to avoid
|
||||
# numerical innacuracies that would otherwise happen
|
||||
# numerical inaccuracies that would otherwise happen
|
||||
# due to downcasting.
|
||||
buffer = torch.zeros((x.size(0), r),
|
||||
dtype=torch.float32,
|
||||
|
||||
@@ -537,7 +537,7 @@ if triton.__version__ >= "2.1.0":
|
||||
alibi_start_q = tl.arange(
|
||||
0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
|
||||
alibi_start_k = cur_batch_ctx_len
|
||||
# # init debuger
|
||||
# # init debugger
|
||||
# offset_db_q = tl.arange(0, BLOCK_M) + block_start_loc
|
||||
# offset_db_k = tl.arange(0, BLOCK_N)
|
||||
# calc q[BLOCK_M, BLOCK_MODEL] mul k[prefix_len: , BLOCK_DMODEL]
|
||||
|
||||
@@ -41,7 +41,7 @@ class DeciLMForCausalLM(LlamaForCausalLM):
|
||||
Based on the llama executor.
|
||||
|
||||
The main difference is that DeciLM uses Variable Grouped Query Attention.
|
||||
The constant number of GQA heads in the decoder is overriden with a value
|
||||
The constant number of GQA heads in the decoder is overridden with a value
|
||||
per layer.
|
||||
|
||||
Usually, in the HuggingFace implementation, instead of
|
||||
|
||||
@@ -36,14 +36,14 @@ def init_custom_ar() -> None:
|
||||
if world_size not in _SUPPORTED_WORLD_SIZES:
|
||||
logger.warn(
|
||||
"Custom allreduce is disabled due to an unsupported world size: "
|
||||
"%d. Supported world sizes: %s. To slience this warning, specify"
|
||||
"%d. Supported world sizes: %s. To silence this warning, specify"
|
||||
"disable_custom_all_reduce=True explicitly.", world_size,
|
||||
str(_SUPPORTED_WORLD_SIZES))
|
||||
return
|
||||
if not _can_p2p(rank, world_size):
|
||||
logger.warn(
|
||||
"Custom allreduce is disabled because your platform lacks GPU P2P"
|
||||
" capability. To slience this warning, specify"
|
||||
" capability. To silence this warning, specify"
|
||||
"disable_custom_all_reduce=True explicitly.")
|
||||
return
|
||||
_CA_HANDLE = CustomAllreduce(rank, world_size)
|
||||
|
||||
@@ -189,7 +189,7 @@ def get_pipeline_model_parallel_next_rank():
|
||||
|
||||
|
||||
def get_pipeline_model_parallel_prev_rank():
|
||||
"""Return the global rank that preceeds the caller in the pipeline"""
|
||||
"""Return the global rank that precedes the caller in the pipeline"""
|
||||
assert _PIPELINE_GLOBAL_RANKS is not None, (
|
||||
"Pipeline parallel group is not initialized")
|
||||
rank_in_pipeline = get_pipeline_model_parallel_rank()
|
||||
|
||||
+1
-1
@@ -204,7 +204,7 @@ def _generate_random_fp8_e5m2(
|
||||
# NOTE(zhaoyang): Due to NaN and Inf representation for fp8 data type,
|
||||
# it may occur Inf or NaN if we directly use torch.randint
|
||||
# to generate random data for fp8 data.
|
||||
# For example, s.11111.00 in fp8e5m2 format repesents Inf.
|
||||
# For example, s.11111.00 in fp8e5m2 format represents Inf.
|
||||
# | E4M3 | E5M2
|
||||
#-----|-------------|-------------------
|
||||
# Inf | N/A | s.11111.00
|
||||
|
||||
Reference in New Issue
Block a user