[don't merge] check mlinter

[ydshieh]

Summary

• Intentionally introduces a TRF008 violation in modeling_kosmos2_5.py (empty @add_start_docstrings("")) to verify that transformers-mlinter runs correctly via make check-code-quality in CI.

Test plan

• CI check_code_quality job should fail on the Modeling file structure step with a TRF008 error

🤖 mlinter review state

• hash: c26b9763
• commit: 1b623ffe
• review: [don't merge] check mlinter #47774 (review)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

aaa

[github-actions]

Model linter — first pass

transformers-mlinter found 5 item(s) in the model files this PR touches. These are structural conventions a maintainer would otherwise flag by hand.

This is automated and advisory — it does not block merging.

rule	count	what it checks
TRF034	2	Layer classes held in an nn.ModuleList must subclass GradientCheckpointingLayer.
TRF029	1	A module taking config must not also take arguments that live on the config.
TRF020	1	MLA models must isolate the KV LoRA expansion (kv_b_proj) in a dedicated method that forward() calls.
TRF008	1	Doc decorators on PreTrainedModel classes should avoid empty add_start_docstrings usage.

4 item(s) outside this diff

• src/transformers/models/axk2/modular_axk2.py:202 TRF029 — TRF029: AXK2GatedRMSNorm.__init__ takes config and also eps. Read those off the config inside the module so there is one source of truth.
• src/transformers/models/axk2/modular_axk2.py:314 TRF020 — TRF020: AXK2Attention declares the KV LoRA expansion self.kv_b_proj but forward does not call a dedicated expansion method. Move the expansion into a method (e.g. expand_kv(k_nope, k_pe) -> key_states, value_states) and call it from forward, so external backends (vLLM/SGLang) can override it and consume the compressed KV cache directly.
• src/transformers/models/granite_swa/modular_granite_swa.py:246 TRF034 — TRF034: GraniteSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.
• src/transformers/models/granitemoe_swa/modular_granitemoe_swa.py:180 TRF034 — TRF034: GraniteMoeSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.

[github-actions]

TRF008 — TRF008: Kosmos2_5Model uses add_start_docstrings without non-empty docstring arguments.

Why this matters

Empty decorator usage produces unclear docs and weakens generated API documentation quality.

-@add_start_docstrings("")
+@add_start_docstrings("The Acme model.")
 class AcmeModel(AcmePreTrainedModel):
     ...

Suppress with # trf-ignore: TRF008 if intentional.

[github-actions]

Model linter — first pass

transformers-mlinter found 5 item(s) in the model files this PR touches. These are structural conventions a maintainer would otherwise flag by hand.

This is automated and advisory — it does not block merging.

rule	count	what it checks
TRF034	2	Layer classes held in an nn.ModuleList must subclass GradientCheckpointingLayer.
TRF029	1	A module taking config must not also take arguments that live on the config.
TRF020	1	MLA models must isolate the KV LoRA expansion (kv_b_proj) in a dedicated method that forward() calls.
TRF008	1	Doc decorators on PreTrainedModel classes should avoid empty add_start_docstrings usage.

4 item(s) outside this diff

• src/transformers/models/axk2/modular_axk2.py:202 TRF029 — TRF029: AXK2GatedRMSNorm.__init__ takes config and also eps. Read those off the config inside the module so there is one source of truth.
• src/transformers/models/axk2/modular_axk2.py:314 TRF020 — TRF020: AXK2Attention declares the KV LoRA expansion self.kv_b_proj but forward does not call a dedicated expansion method. Move the expansion into a method (e.g. expand_kv(k_nope, k_pe) -> key_states, value_states) and call it from forward, so external backends (vLLM/SGLang) can override it and consume the compressed KV cache directly.
• src/transformers/models/granite_swa/modular_granite_swa.py:246 TRF034 — TRF034: GraniteSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.
• src/transformers/models/granitemoe_swa/modular_granitemoe_swa.py:180 TRF034 — TRF034: GraniteMoeSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.

[github-actions]

TRF008 — TRF008: Kosmos2_5Model uses add_start_docstrings without non-empty docstring arguments.

Why this matters

Empty decorator usage produces unclear docs and weakens generated API documentation quality.

-@add_start_docstrings("")
+@add_start_docstrings("The Acme model.")
 class AcmeModel(AcmePreTrainedModel):
     ...

Suppress with # trf-ignore: TRF008 if intentional.

[github-actions]

Model linter — first pass

transformers-mlinter found 5 item(s) in the model files this PR touches. These are structural conventions a maintainer would otherwise flag by hand.

This is automated and advisory — it does not block merging.

rule	count	what it checks
TRF034	2	Layer classes held in an nn.ModuleList must subclass GradientCheckpointingLayer.
TRF029	1	A module taking config must not also take arguments that live on the config.
TRF020	1	MLA models must isolate the KV LoRA expansion (kv_b_proj) in a dedicated method that forward() calls.
TRF008	1	Doc decorators on PreTrainedModel classes should avoid empty add_start_docstrings usage.

4 item(s) outside this diff

• src/transformers/models/axk2/modular_axk2.py:202 TRF029 — TRF029: AXK2GatedRMSNorm.__init__ takes config and also eps. Read those off the config inside the module so there is one source of truth.
• src/transformers/models/axk2/modular_axk2.py:314 TRF020 — TRF020: AXK2Attention declares the KV LoRA expansion self.kv_b_proj but forward does not call a dedicated expansion method. Move the expansion into a method (e.g. expand_kv(k_nope, k_pe) -> key_states, value_states) and call it from forward, so external backends (vLLM/SGLang) can override it and consume the compressed KV cache directly.
• src/transformers/models/granite_swa/modular_granite_swa.py:246 TRF034 — TRF034: GraniteSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.
• src/transformers/models/granitemoe_swa/modular_granitemoe_swa.py:180 TRF034 — TRF034: GraniteMoeSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.

[github-actions]

TRF008 — TRF008: Kosmos2_5Model uses add_start_docstrings without non-empty docstring arguments.

Why this matters

Empty decorator usage produces unclear docs and weakens generated API documentation quality.

-@add_start_docstrings("")
+@add_start_docstrings("The Acme model.")
 class AcmeModel(AcmePreTrainedModel):
     ...

Suppress with # trf-ignore: TRF008 if intentional.

[github-actions]

Model linter — first pass

transformers-mlinter found 5 item(s) in the model files this PR touches. These are structural conventions a maintainer would otherwise flag by hand.

This is automated and advisory — it does not block merging.

rule	count	what it checks
TRF034	2	Layer classes held in an nn.ModuleList must subclass GradientCheckpointingLayer.
TRF029	1	A module taking config must not also take arguments that live on the config.
TRF020	1	MLA models must isolate the KV LoRA expansion (kv_b_proj) in a dedicated method that forward() calls.
TRF008	1	Doc decorators on PreTrainedModel classes should avoid empty add_start_docstrings usage.

4 item(s) outside this diff

• src/transformers/models/axk2/modular_axk2.py:202 TRF029 — TRF029: AXK2GatedRMSNorm.__init__ takes config and also eps. Read those off the config inside the module so there is one source of truth.
• src/transformers/models/axk2/modular_axk2.py:314 TRF020 — TRF020: AXK2Attention declares the KV LoRA expansion self.kv_b_proj but forward does not call a dedicated expansion method. Move the expansion into a method (e.g. expand_kv(k_nope, k_pe) -> key_states, value_states) and call it from forward, so external backends (vLLM/SGLang) can override it and consume the compressed KV cache directly.
• src/transformers/models/granite_swa/modular_granite_swa.py:246 TRF034 — TRF034: GraniteSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.
• src/transformers/models/granitemoe_swa/modular_granitemoe_swa.py:180 TRF034 — TRF034: GraniteMoeSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.

[github-actions]

TRF008 — TRF008: Kosmos2_5Model uses add_start_docstrings without non-empty docstring arguments.

Why this matters

Empty decorator usage produces unclear docs and weakens generated API documentation quality.

-@add_start_docstrings("")
+@add_start_docstrings("The Acme model.")
 class AcmeModel(AcmePreTrainedModel):
     ...

Suppress with # trf-ignore: TRF008 if intentional.

[github-actions]

Model linter — first pass

transformers-mlinter found 6 item(s) in the model files this PR touches. These are structural conventions a maintainer would otherwise flag by hand.

This is automated and advisory — it does not block merging.

rule	count	what it checks
TRF034	2	Layer classes held in an nn.ModuleList must subclass GradientCheckpointingLayer.
TRF004	1	Models must never override tie_weights. Use _tied_weights_keys instead.
TRF029	1	A module taking config must not also take arguments that live on the config.
TRF020	1	MLA models must isolate the KV LoRA expansion (kv_b_proj) in a dedicated method that forward() calls.
TRF008	1	Doc decorators on PreTrainedModel classes should avoid empty add_start_docstrings usage.

4 item(s) outside this diff

• src/transformers/models/axk2/modular_axk2.py:202 TRF029 — TRF029: AXK2GatedRMSNorm.__init__ takes config and also eps. Read those off the config inside the module so there is one source of truth.
• src/transformers/models/axk2/modular_axk2.py:314 TRF020 — TRF020: AXK2Attention declares the KV LoRA expansion self.kv_b_proj but forward does not call a dedicated expansion method. Move the expansion into a method (e.g. expand_kv(k_nope, k_pe) -> key_states, value_states) and call it from forward, so external backends (vLLM/SGLang) can override it and consume the compressed KV cache directly.
• src/transformers/models/granite_swa/modular_granite_swa.py:246 TRF034 — TRF034: GraniteSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.
• src/transformers/models/granitemoe_swa/modular_granitemoe_swa.py:180 TRF034 — TRF034: GraniteMoeSWADecoderLayer is stacked in an nn.ModuleList but does not subclass GradientCheckpointingLayer, so gradient checkpointing silently skips it.

[github-actions]

TRF004 — TRF004: AlbertPreTrainedModel overrides tie_weights. Use _tied_weights_keys class attribute to declare tied weights instead.

Why this matters

Overriding tie_weights leads to bad consequences for loading, device_map computation, and saving. Use _tied_weights_keys class attribute to declare tied weights instead.

-def tie_weights(self):
-    self.lm_head.weight = self.emb.weight
+class AcmeForCausalLM(AcmePreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]

Suppress with # trf-ignore: TRF004 if intentional.

[github-actions]

TRF008 — TRF008: Kosmos2_5Model uses add_start_docstrings without non-empty docstring arguments.

Why this matters

Empty decorator usage produces unclear docs and weakens generated API documentation quality.

-@add_start_docstrings("")
+@add_start_docstrings("The Acme model.")
 class AcmeModel(AcmePreTrainedModel):
     ...

Suppress with # trf-ignore: TRF008 if intentional.

[github-actions]

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

run-slow: albert, kosmos2_5

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 31180676409:1
Result: failure | Jobs: 2 | Tests: 10 | Failures: 2 | Duration: 49s

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