[Bugfix][Model] Remove SciPy dependency from Inkling scale planning (#49485)

Signed-off-by: Michael Gschwind <mgschwind@nvidia.com>
Co-authored-by: Michael Gschwind <mgschwind@nvidia.com>
Co-authored-by: OpenAI Codex <codex@openai.com>

tests/models/inkling/test_contract_validation.py

@@ -6,6 +6,7 @@ import pytest
 
 from vllm.config.compilation import CompilationConfig, CUDAGraphMode
 from vllm.models.inkling.common.mm_preprocess import InklingMultiModalDataParser
+from vllm.models.inkling.common.towers import plan_out_scales
 from vllm.models.inkling.configs import (
     InklingAudioConfig,
     InklingModelConfig,
@@ -17,6 +18,22 @@ from vllm.models.inkling.nvidia.sconv_swa_attn import (
 from vllm.v1.attention.backend import AttentionCGSupport
 
 
+def test_vision_scale_plan_matches_released_config():
+    assert plan_out_scales(2, 40, 4) == [
+        (1, 1, 1, 3),
+        (1, 5, 5, 128),
+        (1, 10, 10, 320),
+        (1, 40, 40, 4800),
+        (2, 40, 40, 9600),
+    ]
+
+
+def test_vision_scale_plan_breaks_assignment_ties_in_order():
+    reductions = [np.prod(scale[:-1]) for scale in plan_out_scales(2, 52, 4)]
+
+    assert reductions == sorted(set(reductions))
+
+
 @pytest.mark.parametrize(
     ("config_cls", "kwargs", "missing"),
     [

vllm/models/inkling/common/towers.py

@@ -9,6 +9,7 @@ Both use vLLM's standard ``RMSNorm`` (CPU-friendly, with a native fallback).
 
 from __future__ import annotations
 
+from itertools import combinations
 from typing import cast
 
 import numpy as np
@@ -45,6 +46,20 @@ def _prime_factors(n: int) -> list[int]:
     return factors
 
 
+def linear_sum_assignment(
+    cost_matrix: np.ndarray,
+) -> tuple[np.ndarray, np.ndarray]:
+    """Implement SciPy's assignment for Inkling's ordered L1 cost matrix."""
+    rows = np.arange(cost_matrix.shape[0])
+    cols = np.array(
+        min(
+            combinations(range(cost_matrix.shape[1]), len(rows)),
+            key=lambda candidate: cost_matrix[rows, candidate].sum(),
+        )
+    )
+    return rows, cols
+
+
 def plan_out_scales(
     temporal_patch_size: int, patch_size: int, n_layers: int, n_channels: int = 3
 ) -> list[tuple[int, int, int, int]]:
@@ -97,8 +112,6 @@ def plan_out_scales(
     if n_layers >= len(scales):
         idxs = np.argmin(cost_matrix, axis=1)
     else:
-        from scipy.optimize import linear_sum_assignment
-
         idxs = linear_sum_assignment(cost_matrix)[1]
 
     assert len(idxs) >= 2