fix: make num_experts BC fallback not override explicit n_routed_experts

[webtecnica]

Description

Fixes #47355

Root cause: In both GlmMoeDsaConfig.__post_init__ and DeepseekV32Config.__post_init__, the backward-compatibility re-route from num_experts (legacy key) to n_routed_experts was unconditional — it could overwrite an explicit n_routed_experts value from the same config.json.

Example: A config with {"n_routed_experts": 168, "num_experts": 256} would silently use 256 experts, causing an OOM downstream (vLLM allocates for 256 experts instead of 168).

Fix: Only apply the BC fallback when n_routed_experts is still at its defining class's default value, which means it was NOT explicitly provided. Since n_routed_experts is a dataclass field consumed by __init__ before __post_init__ runs, we compare self.n_routed_experts against type(self).n_routed_experts.

Files changed:

File	Change
modular_deepseek_v32.py	Root fix — source that regenerates generated files
configuration_deepseek_v32.py	Generated copy of the same fix
configuration_glm_moe_dsa.py	Generated copy (unwrapped from modular)
tests/models/glm_moe_dsa/test_config_glm_moe_dsa.py	3 test cases covering the fix + BC preservation

Verification (6 test cases, all pass):

Scenario	Expected	Result
Both keys provided	n_routed_experts=168 wins	✅
Only legacy num_experts	BC fallback applies: 128	✅
Only modern n_routed_experts	Preserved: 42	✅
DeepseekV32Config parent	Same fix: 64	✅
Both keys same value (256)	No issue	✅
Not provided either	Class default (256)	✅

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_v32, glm_moe_dsa

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 29672193811
Result: failure | Grafana metrics are not available yet.

Code quality check failed: test jobs were skipped. Fix the code quality issues and push again to run tests.