Skip to content

[Won't Merge]Qwen fp8 #2196

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

Closed
yiliu30 wants to merge 5 commits into from
Closed

[Won't Merge]Qwen fp8 #2196

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5f61128
update moe
yiliu30 Apr 28, 2025
c3bd0f5
remoce unused code
Yi4Liu Apr 29, 2025
53a76b7
update moe experts
Yi4Liu Apr 29, 2025
9dd74f9
revert change
Yi4Liu Apr 29, 2025
ce9fc55
remove const
Yi4Liu Jun 4, 2025
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
21 changes: 0 additions & 21 deletions neural_compressor/common/utils/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,27 +21,6 @@
from neural_compressor.common.utils.utility import *


# FIXME: (Yi) REMOVE BELOW CODE
import os

DEEPSEEK_EXPERTS = 256
VLLM_TP_SIZE = int(os.getenv("VLLM_TP_SIZE", "8"))
VLLM_EP_SIZE = int(os.getenv("VLLM_EP_SIZE", VLLM_TP_SIZE))
NUM_EXPERTS_PER_EP_RANK = DEEPSEEK_EXPERTS // VLLM_EP_SIZE # 32
VLLM_MOE_N_SLICE = int(os.getenv("VLLM_MOE_N_SLICE", 8))
NUM_EXPERTS_PER_GROUP_PER_RANK = NUM_EXPERTS_PER_EP_RANK // VLLM_MOE_N_SLICE # 4
FUSED_MOE_EXPERTS = NUM_EXPERTS_PER_GROUP_PER_RANK # 4

logger.warning_once(
(
f"INC uses VLLM_TP_SIZE={VLLM_TP_SIZE},\n"
f"VLLM_EP_SIZE={VLLM_EP_SIZE},\n"
f"NUM_EXPERTS_PER_EP_RANK={NUM_EXPERTS_PER_EP_RANK},\n"
f"VLLM_MOE_N_SLICE={VLLM_MOE_N_SLICE},\n"
f"NUM_EXPERTS_PER_GROUP_PER_RANK={NUM_EXPERTS_PER_GROUP_PER_RANK},\n"
f"FUSED_MOE_EXPERTS={FUSED_MOE_EXPERTS}"
)
)

import sys
import pdb
Expand Down
15 changes: 7 additions & 8 deletions neural_compressor/torch/algorithms/fp8_quant/_core/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,6 @@
get_patched_module_table,
get_patched_module_type_table,
)
from neural_compressor.common import utils as inc_utils
from neural_compressor.torch.utils.auto_accelerator import auto_detect_accelerator, INCAcceleratorType
deepspeed_exists = False
if importlib.util.find_spec("deepspeed"): # check if deepspeed is installed
Expand All @@ -59,7 +58,7 @@ def maybe_dequant_original_fp8_weight(mod: torch.nn.Module, param: torch.Tensor)
"dynamic_moe": ModuleType(
1,
[],
inc_utils.FUSED_MOE_EXPERTS + 1, # FIXME (Yi) # one output, FUSED_MOE_EXPERTS weights
8 + 1,
True,
),
}
Expand Down Expand Up @@ -235,12 +234,12 @@ def convert_scales_to_tensors_dict(scales_obj, scales_file_format, hp_dtype, dev
"VllmMixtureOfExpertsOpFP8": ModuleInfo("dynamic_moe", PatchedVllmMixtureOfExpertsOpFP8),
# FIXME (Yi) revert change
"FusedMoE": ModuleInfo("linear", PatchedMixtralMoE, False),
# "GaudiMixtralSparseMoeBlock": ModuleInfo("dynamic_moe", PatchedGaudiMixtralSparseMoeBlock),
# "VllmMixtureOfExpertsOp": (
# ModuleInfo("dynamic_moe", PatchedVllmMixtureOfExpertsOpV2)
# if os.getenv("LOW_CPU_MEM", "0") == "1"
# else ModuleInfo("dynamic_moe", PatchedVllmMixtureOfExpertsOpV1)
# ),
"GaudiMixtralSparseMoeBlock": ModuleInfo("dynamic_moe", PatchedGaudiMixtralSparseMoeBlock),
"VllmMixtureOfExpertsOp": (
ModuleInfo("dynamic_moe", PatchedVllmMixtureOfExpertsOpV2)
if os.getenv("LOW_CPU_MEM", "0") == "1"
else ModuleInfo("dynamic_moe", PatchedVllmMixtureOfExpertsOpV1)
),
}


Expand Down
2 changes: 2 additions & 0 deletions neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -781,6 +781,7 @@ def forward_quant(self,
hidden_states,
expert_routing_table,
router_weights,
layer=None,
permuted_weights=True,
activation="silu"):
experts_range = range(self.num_experts)
Expand Down Expand Up @@ -810,6 +811,7 @@ def forward_measure(self,
hidden_states,
expert_routing_table,
router_weights,
layer=None,
permuted_weights=True,
activation="silu"):
experts_range = range(self.num_experts)
Expand Down