Mamba MOE Quant Configs + Fix Export Bug (#882)

## What does this PR do?

**Type of change:** ? Bug fix

**Overview:** ?
- Fix a bug in MCore export `exclude_modules` where the layers had an
extra period at the end
- Add custom quant configs for mamba moes

## Usage
<!-- You can potentially add a usage example below. -->

```python
# Add a code snippet demonstrating how to use this
```

## Testing
<!-- Mention how have you tested your change if applicable. -->

## Before your PR is "*Ready for review*"
<!-- If you haven't finished some of the above items you can still open
`Draft` PR. -->

- **Make sure you read and follow [Contributor
guidelines](https://github.com/NVIDIA/Model-Optimizer/blob/main/CONTRIBUTING.md)**
and your commits are signed.
- **Is this change backward compatible?**: Yes/No <!--- If No, explain
why. -->
- **Did you write any new necessary tests?**: Yes/No
- **Did you add or update any necessary documentation?**: Yes/No
- **Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?**:
Yes/No <!--- Only for new features, API changes, critical bug fixes or
bw breaking changes. -->

## Additional Information
<!-- E.g. related issue. -->

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **New Features**
* Added four new Mamba MOE quantization configurations: aggressive and
conservative variants for both FP8 and NVFP4 quantization schemes,
providing enhanced flexibility in quantization options for different use
cases.

* **Bug Fixes**
* Improved quantization export module exclusion pattern handling to
properly normalize trailing dots from exclude patterns during export.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
Signed-off-by: Daniel Korzekwa <dkorzekwa@nvidia.com>

Author: jenchen13

modelopt/torch/quantization/calib/max.py

@@ -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:

modelopt/torch/quantization/config.py

@@ -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},
@@ -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},
@@ -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": {
@@ -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"]