apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-27 04:46:16 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

deepgemm pre-compile tool support mixed parallel (#4282)
Some checks are pending
CE Compile Job / ce_job_pre_check (push) Waiting to run
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Blocked by required conditions
Details
CE Compile Job / FD-Clone-Linux (push) Blocked by required conditions
Details
CE Compile Job / Show Code Archive Output (push) Blocked by required conditions
Details
CE Compile Job / BUILD_SM8090 (push) Blocked by required conditions
Details
CE Compile Job / BUILD_SM8689 (push) Blocked by required conditions
Details
CE Compile Job / CE_UPLOAD (push) Blocked by required conditions
Details
Deploy GitHub Pages / deploy (push) Waiting to run
Details

Browse Source
This commit is contained in:
GoldPancake
2025-09-26 18:43:39 +08:00
committed by GitHub
parent 5c6e859681
commit 79c2c52756
1 changed files with 27 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
tools/deep_gemm_pre-compile/generate_config.py
View File

@@ -41,51 +41,26 @@ def generate_kn_pairs(args, model_cfg: dict) -> Tuple[List, List, List]:
gemm_kn_pairs = []
grouped_gemm_contiguous_kn_pairs = []
grouped_gemm_masked_kn_pairs = []
if tp_size > 1 and ep_size == 1:
logger.debug("Generating kn pairs for tensor parallel.")
# Dense normal gemm
gemm_kn_pairs.extend(
[
[int(intermediate_size / tp_size), hidden_size],
[hidden_size, int(head_dim * (num_attention_heads + num_key_value_heads * 2) / tp_size)],
[hidden_size, int(intermediate_size * 2 / tp_size)],
[int(hidden_size / tp_size), hidden_size],
]
)
logger.debug("Generating kn pairs for tensor parallel.")
# Dense normal gemm
gemm_kn_pairs.extend(
[
[int(intermediate_size / tp_size), hidden_size],
[hidden_size, int(head_dim * (num_attention_heads + num_key_value_heads * 2) / tp_size)],
[hidden_size, int(intermediate_size * 2 / tp_size)],
[int(hidden_size / tp_size), hidden_size],
]
)
# Moe grouped gemm contiguous
grouped_gemm_contiguous_kn_pairs.extend(
[
[int(moe_intermediate_size / tp_size), hidden_size],
[hidden_size, int(moe_intermediate_size * 2 / tp_size)],
]
)
if has_shared_experts:
logger.debug("Generating kn pairs for models with shared experts.")
gemm_kn_pairs.extend(
[
[hidden_size, int(moe_intermediate_size * 4 / tp_size)],
[int(moe_intermediate_size * 2 / tp_size), hidden_size],
]
)
elif tp_size == 1 and ep_size > 1:
logger.debug("Generating kn pairs for expert parallel.")
# Dense normal gemm
gemm_kn_pairs.extend(
[
[intermediate_size, hidden_size],
[hidden_size, int(head_dim * (num_attention_heads + num_key_value_heads * 2))],
[hidden_size, int(intermediate_size * 2)],
[hidden_size, hidden_size],
]
)
# Moe grouped gemm contiguous
grouped_gemm_contiguous_kn_pairs.extend(
[
[moe_intermediate_size, hidden_size],
[hidden_size, int(moe_intermediate_size * 2)],
]
)
# Moe grouped gemm contiguous
grouped_gemm_contiguous_kn_pairs.extend(
[
[int(moe_intermediate_size / tp_size), hidden_size],
[hidden_size, int(moe_intermediate_size * 2 / tp_size)],
]
)
if ep_size > 1:
# Moe grouped gemm masked
grouped_gemm_masked_kn_pairs.extend(
[
@@ -93,18 +68,14 @@ def generate_kn_pairs(args, model_cfg: dict) -> Tuple[List, List, List]:
[hidden_size, int(moe_intermediate_size * 2)],
]
)
if has_shared_experts:
logger.debug("Generating kn pairs for models with shared experts.")
gemm_kn_pairs.extend(
[
[hidden_size, int(moe_intermediate_size * 4)],
[int(moe_intermediate_size * 2), hidden_size],
]
)
elif tp_size > 1 and ep_size > 1:
raise ValueError("Not supported to enable EP and TP at the same time for now.")
else:
raise ValueError("Please check the tensor parallel size and expert parallel size.")
if has_shared_experts:
logger.debug("Generating kn pairs for models with shared experts.")
gemm_kn_pairs.extend(
[
[hidden_size, int(moe_intermediate_size * 4 / tp_size)],
[int(moe_intermediate_size * 2 / tp_size), hidden_size],
]
)
return (
gemm_kn_pairs,