Skip to content

Mamba MOE Quant Configs + Fix Export Bug #882

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

Merged
jenchen13 merged 4 commits into from
Feb 17, 2026
+82 −4
Merged

Mamba MOE Quant Configs + Fix Export Bug #882

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
5bf3432
mamba configs
jenchen13 Feb 12, 2026
4d0f1f6
change assert order
jenchen13 Feb 12, 2026
9fbe89c
add o_proj
jenchen13 Feb 13, 2026
c9fb020
lint
jenchen13 Feb 13, 2026
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
6 changes: 3 additions & 3 deletions modelopt/torch/quantization/calib/max.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,15 +66,15 @@ def collect(self, x):
if x.device.type == "meta":
self._calib_amax = local_amax
return
assert not torch.any(torch.isnan(local_amax)), (
f"detected nan values in amax. nan in original tensor: {torch.any(torch.isnan(x))}"
)
assert torch.all(local_amax >= 0), (
"detected negative values after abs, could be torch or cuda bug"
)
assert not torch.any(torch.isinf(local_amax)), (
f"detected inf values in amax. inf in original tensor: {torch.any(torch.isinf(x))}"
)
assert not torch.any(torch.isnan(local_amax)), (
f"detected nan values in amax. nan in original tensor: {torch.any(torch.isnan(x))}"
)
if self._calib_amax is None:
self._calib_amax = local_amax
else:
Expand Down
80 changes: 79 additions & 1 deletion modelopt/torch/quantization/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,12 +156,21 @@
"*mlp.gate.*": {"enable": False}, # Skip the MOE router
"*mlp.shared_expert_gate.*": {"enable": False}, # Skip the MOE router
"*linear_attn.conv1d*": {"enable": False},
"*mixer.conv1d*": {"enable": False},
"*mixer.conv1d*": {"enable": False}, # Skip mamba conv1d
"*output_layer*": {"enable": False},
"output.*": {"enable": False},
"default": {"enable": False},
}

_mamba_moe_disabled_quantizer_cfg = {
"*fc1_latent_proj*": {"enable": False}, # Skip Latent MOE
"*fc2_latent_proj*": {"enable": False}, # Skip Latent MOE
"*q_proj*": {"enable": False}, # Skip QKV Linear
"*k_proj*": {"enable": False}, # Skip QKV Linear
"*v_proj*": {"enable": False}, # Skip QKV Linear
"*o_proj*": {"enable": False}, # Skip QKV Output Projection
}

INT8_DEFAULT_CFG = {
"quant_cfg": {
"*weight_quantizer": {"num_bits": 8, "axis": 0},
Expand Down Expand Up @@ -198,6 +207,28 @@
"algorithm": "max",
}

MAMBA_MOE_FP8_AGGRESSIVE_CFG = {
"quant_cfg": {
"*weight_quantizer": {"num_bits": (4, 3), "axis": None},
"*input_quantizer": {"num_bits": (4, 3), "axis": None},
**_default_disabled_quantizer_cfg,
**_mamba_moe_disabled_quantizer_cfg,
},
"algorithm": "max",
}

MAMBA_MOE_FP8_CONSERVATIVE_CFG = {
"quant_cfg": {
"*weight_quantizer": {"num_bits": (4, 3), "axis": None},
"*input_quantizer": {"num_bits": (4, 3), "axis": None},
**_default_disabled_quantizer_cfg,
**_mamba_moe_disabled_quantizer_cfg,
"*mixer.in_proj*": {"enable": False}, # Skip mamba linear
"*mixer.out_proj*": {"enable": False}, # Skip mamba linear
},
"algorithm": "max",
}

FP8_PER_CHANNEL_PER_TOKEN_CFG = {
"quant_cfg": {
"*weight_quantizer": {"num_bits": (4, 3), "axis": 0},
Expand Down Expand Up @@ -388,6 +419,49 @@
"algorithm": "max",
}


MAMBA_MOE_NVFP4_AGGRESSIVE_CFG = {
"quant_cfg": {
"*weight_quantizer": {
"num_bits": (2, 1),
"block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
"axis": None,
"enable": True,
},
"*input_quantizer": {
"num_bits": (2, 1),
"block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
"axis": None,
"enable": True,
},
**_default_disabled_quantizer_cfg,
**_mamba_moe_disabled_quantizer_cfg,
},
"algorithm": "max",
}
MAMBA_MOE_NVFP4_CONSERVATIVE_CFG = {
"quant_cfg": {
"*weight_quantizer": {
"num_bits": (2, 1),
"block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
"axis": None,
"enable": True,
},
"*input_quantizer": {
"num_bits": (2, 1),
"block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
"axis": None,
"enable": True,
},
**_default_disabled_quantizer_cfg,
**_mamba_moe_disabled_quantizer_cfg,
"*mixer.in_proj*": {"enable": False}, # Skip mamba linear
"*mixer.out_proj*": {"enable": False}, # Skip mamba linear
},
"algorithm": "max",
}


NVFP4_AWQ_LITE_CFG = {
"quant_cfg": {
"*weight_quantizer": {
Expand Down Expand Up @@ -652,6 +726,10 @@
"NVFP4_MLP_WEIGHT_ONLY_CFG",
"MXFP4_MLP_WEIGHT_ONLY_CFG",
"NVFP4_MLP_ONLY_CFG",
"MAMBA_MOE_NVFP4_CONSERVATIVE_CFG",
"MAMBA_MOE_NVFP4_AGGRESSIVE_CFG",
"MAMBA_MOE_FP8_CONSERVATIVE_CFG",
"MAMBA_MOE_FP8_AGGRESSIVE_CFG",
}

BiasType = Literal["static", "dynamic"]
Expand Down