Phase B + C: Qwen3.5 own-the-stack — fork-owned model class + layers

.gitignore

@@ -253,6 +253,10 @@ benchmarks/nvllm/results/*.sqlite
 # nvllm evaluation results (evals/)
 benchmarks/nvllm/results/evals/
 
+# nvllm trace full-logs (keep trimmed decode_log.txt tracked, full dumps local-only)
+benchmarks/nvllm/traces/**/*.full.txt
+benchmarks/nvllm/traces/**/*.full.log
+
 # nvllm: excluded from fork
 CLAUDE.md
 .claude/

benchmarks/nvllm/traces/cute_fusion/2026-04-17-phase-c/decode_log.txt

@@ -0,0 +1,32 @@
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.3.self_attn.attn nat=1 phaseB  ref: absmax=220.6929 mean=5.4770e-02  kernel: absmax=220.6929 mean=5.4770e-02  diff: max=0.0000 mean=2.0900e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.3.self_attn.attn nat=1 phaseC  hidden_ref_absmax=0.0000 hidden_kernel_absmax=0.0000 h_max_diff=0.0000  res_ref_absmax=170141183460469231731687303715884105728.0000 res_kernel_absmax=170141183460469231731687303715884105728.0000 r_max_diff=224.0000  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.7.self_attn.attn nat=1 phaseB  ref: absmax=50.6014 mean=1.1364e-02  kernel: absmax=50.6014 mean=1.1364e-02  diff: max=0.0000 mean=1.4823e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.7.self_attn.attn nat=1 phaseC  hidden_ref_absmax=65.5932 hidden_kernel_absmax=65.5000 h_max_diff=0.0932  res_ref_absmax=50.6014 res_kernel_absmax=50.5000 r_max_diff=0.1014  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.11.self_attn.attn nat=1 phaseB  ref: absmax=39.5711 mean=1.0581e-03  kernel: absmax=39.5711 mean=1.0581e-03  diff: max=0.0000 mean=2.0968e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.11.self_attn.attn nat=1 phaseC  hidden_ref_absmax=53.4998 hidden_kernel_absmax=53.5000 h_max_diff=0.0063  res_ref_absmax=39.5711 res_kernel_absmax=39.5000 r_max_diff=0.0711  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.15.self_attn.attn nat=1 phaseB  ref: absmax=13.8859 mean=3.0576e-03  kernel: absmax=13.8859 mean=3.0576e-03  diff: max=0.0000 mean=1.3743e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.15.self_attn.attn nat=1 phaseC  hidden_ref_absmax=36.4917 hidden_kernel_absmax=36.5000 h_max_diff=0.0083  res_ref_absmax=13.8859 res_kernel_absmax=13.8750 r_max_diff=0.0109  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.19.self_attn.attn nat=1 phaseB  ref: absmax=3.6525 mean=1.5460e-03  kernel: absmax=3.6525 mean=1.5460e-03  diff: max=0.0000 mean=1.5882e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.19.self_attn.attn nat=1 phaseC  hidden_ref_absmax=9.1445 hidden_kernel_absmax=9.1250 h_max_diff=0.0195  res_ref_absmax=3.6525 res_kernel_absmax=3.6562 r_max_diff=0.0038  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.23.self_attn.attn nat=1 phaseB  ref: absmax=10.7538 mean=-4.6950e-04  kernel: absmax=10.7538 mean=-4.6950e-04  diff: max=0.0000 mean=1.8470e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.23.self_attn.attn nat=1 phaseC  hidden_ref_absmax=19.5012 hidden_kernel_absmax=19.5000 h_max_diff=0.0073  res_ref_absmax=10.7538 res_kernel_absmax=10.7500 r_max_diff=0.0065  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.27.self_attn.attn nat=1 phaseB  ref: absmax=3.7602 mean=-7.8647e-04  kernel: absmax=3.7602 mean=-7.8647e-04  diff: max=0.0000 mean=2.5609e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.27.self_attn.attn nat=1 phaseC  hidden_ref_absmax=5.6586 hidden_kernel_absmax=5.6562 h_max_diff=0.0039  res_ref_absmax=3.7602 res_kernel_absmax=3.7656 r_max_diff=0.0070  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.31.self_attn.attn nat=1 phaseB  ref: absmax=7.0185 mean=-4.4733e-03  kernel: absmax=7.0185 mean=-4.4733e-03  diff: max=0.0000 mean=2.2128e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.31.self_attn.attn nat=1 phaseC  hidden_ref_absmax=10.2285 hidden_kernel_absmax=10.2500 h_max_diff=0.0215  res_ref_absmax=7.0185 res_kernel_absmax=7.0312 r_max_diff=0.0128  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.35.self_attn.attn nat=1 phaseB  ref: absmax=3.9329 mean=2.2475e-04  kernel: absmax=3.9329 mean=2.2475e-04  diff: max=0.0000 mean=2.4080e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.35.self_attn.attn nat=1 phaseC  hidden_ref_absmax=4.7694 hidden_kernel_absmax=4.7812 h_max_diff=0.0118  res_ref_absmax=3.9329 res_kernel_absmax=3.9375 r_max_diff=0.0065  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.39.self_attn.attn nat=1 phaseB  ref: absmax=2.0009 mean=-7.4135e-03  kernel: absmax=2.0009 mean=-7.4135e-03  diff: max=0.0000 mean=2.2953e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:36 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.39.self_attn.attn nat=1 phaseC  hidden_ref_absmax=1.7291 hidden_kernel_absmax=1.7266 h_max_diff=0.0034  res_ref_absmax=2.0009 res_kernel_absmax=2.0000 r_max_diff=0.0039  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.43.self_attn.attn nat=1 phaseB  ref: absmax=7.0806 mean=1.8172e-03  kernel: absmax=7.0806 mean=1.8172e-03  diff: max=0.0000 mean=3.5171e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.43.self_attn.attn nat=1 phaseC  hidden_ref_absmax=2.4871 hidden_kernel_absmax=2.4844 h_max_diff=0.0039  res_ref_absmax=7.0806 res_kernel_absmax=7.0938 r_max_diff=0.0131  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.47.self_attn.attn nat=1 phaseB  ref: absmax=11.1461 mean=2.6451e-04  kernel: absmax=11.1461 mean=2.6450e-04  diff: max=0.0000 mean=3.2844e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.47.self_attn.attn nat=1 phaseC  hidden_ref_absmax=8.4472 hidden_kernel_absmax=8.4375 h_max_diff=0.0097  res_ref_absmax=11.1461 res_kernel_absmax=11.1250 r_max_diff=0.0211  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.51.self_attn.attn nat=1 phaseB  ref: absmax=9.8056 mean=-1.3065e-03  kernel: absmax=9.8056 mean=-1.3065e-03  diff: max=0.0000 mean=4.1282e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.51.self_attn.attn nat=1 phaseC  hidden_ref_absmax=4.8361 hidden_kernel_absmax=4.8438 h_max_diff=0.0076  res_ref_absmax=9.8056 res_kernel_absmax=9.8125 r_max_diff=0.0154  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.55.self_attn.attn nat=1 phaseB  ref: absmax=35.0521 mean=-2.5966e-02  kernel: absmax=35.0521 mean=-2.5966e-02  diff: max=0.0000 mean=3.9317e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.55.self_attn.attn nat=1 phaseC  hidden_ref_absmax=19.2748 hidden_kernel_absmax=19.2500 h_max_diff=0.0248  res_ref_absmax=35.0521 res_kernel_absmax=35.0000 r_max_diff=0.0521  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.59.self_attn.attn nat=1 phaseB  ref: absmax=53.3233 mean=-8.7635e-02  kernel: absmax=53.3234 mean=-8.7635e-02  diff: max=0.0000 mean=8.4114e-07  close=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.59.self_attn.attn nat=1 phaseC  hidden_ref_absmax=7.9966 hidden_kernel_absmax=8.0000 h_max_diff=0.0037  res_ref_absmax=53.3234 res_kernel_absmax=53.2500 r_max_diff=0.0734  close_h=True close_r=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:578] [CUTE_DEBUG_FUSION] layer=model.layers.63.self_attn.attn nat=1 phaseB  ref: absmax=75.9949 mean=-1.0771e-01  kernel: absmax=75.9949 mean=-1.0771e-01  diff: max=0.0000 mean=2.0806e-06  close=True
+(EngineCore pid=144) INFO 04-17 20:20:37 [_backend.py:607] [CUTE_DEBUG_FUSION] layer=model.layers.63.self_attn.attn nat=1 phaseC  hidden_ref_absmax=1.2595 hidden_kernel_absmax=1.2578 h_max_diff=0.0030  res_ref_absmax=75.9949 res_kernel_absmax=76.0000 r_max_diff=0.1223  close_h=True close_r=True

benchmarks/nvllm/traces/cute_fusion/2026-04-17-phase-c/summary.md

@@ -0,0 +1,86 @@
+# Phase C ship-gate evidence
+
+**Date:** 2026-04-17
+**Commit:** `cbfadb6a9` (`cbfadb6a9104ea49da1c5a8cbd118cb157662978`)
+**Branch:** `feat/own-the-stack-phase-c`
+**Image:** `nvllm:gb10-ots-phaseC` (ID `6b6c71c9daa3`, 20GB)
+**Model:** `natfii/Qwen3.5-27B-NVFP4-Opus-GB10`
+**Backend:** CuTe paged, PIECEWISE CUDA graphs
+**KV cache:** fp8_e4m3
+**Context:** 65536 tokens, max_num_seqs=4
+**Env:** `CUTE_DEBUG_FUSION=1`
+
+## Ship gate
+
+GSM8K sanity: **8/8 correct (100%) — PASS**
+
+```
+  Q1: OK  expected=72  got=72  (36.6s — first-request warmup)
+  Q2: OK  expected=10  got=10  (6.1s)
+  Q3: OK  expected=5   got=5   (6.4s)
+  Q4: OK  expected=42  got=42  (6.3s)
+  Q5: OK  expected=624 got=624 (6.1s)
+  Q6: OK  expected=35  got=35  (6.4s)
+  Q7: OK  expected=48  got=48  (6.2s)
+  Q8: OK  expected=16  got=16  (6.2s)
+```
+
+Matches Phase B baseline (commit `4110dc77a`, image `nvllm:gb10-ots`) — same 8/8, same ~6 s warm latency per prompt after first-request JIT warmup.
+
+## Tier-1 host-side evidence
+
+`notebooks/nvllm/layers_smoke_tests.py` — **5/5 passed** on host CPU:
+- `test_rms_block_level_equivalence` — `Qwen3_5RMSNorm(5120)` vs `GemmaRMSNorm(5120)`, `rtol=0 atol=0` on bf16 input.
+- `test_rms_head_dim_equivalence` — same at `hidden_size=256` (q/k_norm shape).
+- `test_rms_fused_residual_equivalence` — fused add+norm path.
+- `test_customop_registry_no_collision` — `qwen3_5_rms_norm` and `gemma_rms_norm` coexist in `op_registry`.
+- `test_mlp_class_surface` — `Qwen3_5MLP.__init__` has exactly `(hidden_size, intermediate_size, hidden_act, quant_config, prefix)` — no `expert_gate`, no `reduce_results`.
+
+## Tier-2 lint evidence
+
+- `pre-commit run --files ...` — all hooks pass (ruff, format, typos, SPDX, etc.).
+  SPDX-header hook auto-added canonical vLLM-project copyright line to the 4 new files during first run; re-run clean.
+- `pre-commit run mypy-3.10 --all-files --hook-stage manual` — zero errors in Phase C files.
+  11 pre-existing errors remain in `vllm/v1/attention/backends/cute_paged/{kernel.py,_backend.py}` and `vllm/v1/core/kv_cache_utils.py` — identical to Phase B baseline, not introduced by this refactor.
+
+## Tier-3 image-content verification
+
+```bash
+docker run --rm --gpus all --entrypoint python nvllm:gb10-ots-phaseC \
+  -c "from vllm.nvllm.layers import Qwen3_5RMSNorm, Qwen3_5MLP; \
+      from vllm.model_executor.custom_op import op_registry; \
+      print('IMPORT_OK'); \
+      print('qwen3_5_rms_norm in registry:', 'qwen3_5_rms_norm' in op_registry)"
+# → IMPORT_OK
+# → qwen3_5_rms_norm in registry: True
+```
+
+## Repro
+
+```bash
+git checkout feat/own-the-stack-phase-c
+docker build -f docker/Dockerfile.gb10 -t nvllm:gb10-ots-phaseC .
+NVLLM_IMAGE=nvllm:gb10-ots-phaseC CUTE_DEBUG_FUSION=1 ./scripts/serve-cute.sh
+# wait for "Application startup complete."
+.venv/bin/python scripts/gsm8k_sanity.py --api http://localhost:8000/v1 --model default
+```
+
+## Rollback
+
+Phase B image `nvllm:gb10-ots` stays on disk as the rollback snapshot.
+`nvllm:gb10-preshim-phaseC-20260417` tag also points at the same Phase B image ID (`240c48497d10`).
+
+```bash
+docker tag nvllm:gb10-ots nvllm:gb10
+git checkout main
+```
+
+No data migration, no weight-format changes, no config-schema changes — rollback is instant.
+
+## Kernel durations
+
+No kernel changes in Phase C — this is a pure layer-ownership refactor.
+Decode kernel durations match the Phase B baseline
+(see [`../2026-04-17-own-the-stack/summary.md`](../2026-04-17-own-the-stack/summary.md)).
+No new `.nsys-rep` required per AGENTS.md §4 — new traces are only mandatory for perf claims,
+not for semantically-equivalent refactors.

notebooks/nvllm/layers_smoke_tests.py

@@ -0,0 +1,148 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the nvllm fork
+"""Tier-1 host-side smoke tests for nvllm Phase C owned layer classes.
+
+Run:  .venv/bin/python notebooks/nvllm/layers_smoke_tests.py
+Runs in ~5 s on CPU. No GPU required.
+
+Tests:
+  1. Qwen3_5RMSNorm block-level equivalence vs upstream GemmaRMSNorm(5120).
+  2. Qwen3_5RMSNorm head-dim equivalence vs upstream GemmaRMSNorm(256).
+  3. Qwen3_5RMSNorm fused-residual equivalence vs upstream GemmaRMSNorm.
+  4. CustomOp registry: 'qwen3_5_rms_norm' exists and does not clobber
+     'gemma_rms_norm'.
+
+Notes:
+  - Qwen3_5MLP vs Qwen2MoeMLP equivalence test is SKIPPED here because
+    MergedColumnParallelLinear / RowParallelLinear require an initialized
+    distributed backend (torch.distributed). Shape-only smoke test is
+    included. Full correctness is covered by Tier-3 GSM8K 8/8.
+"""
+
+import sys
+
+import torch
+
+from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.model_executor.custom_op import op_registry
+from vllm.model_executor.layers.layernorm import GemmaRMSNorm
+from vllm.nvllm.layers.layernorm import Qwen3_5RMSNorm
+from vllm.nvllm.layers.mlp import Qwen3_5MLP
+
+
+def _seed_weights(rms_upstream: GemmaRMSNorm, rms_owned: Qwen3_5RMSNorm) -> None:
+    """Copy identical weights into both norms so outputs can be compared."""
+    torch.manual_seed(42)
+    w = torch.randn_like(rms_upstream.weight.data)
+    rms_upstream.weight.data.copy_(w)
+    rms_owned.weight.data.copy_(w)
+
+
+def test_rms_block_level_equivalence() -> None:
+    """Block-level (hidden_size=5120) equivalence — no residual."""
+    hidden = 5120
+    upstream = GemmaRMSNorm(hidden, eps=1e-6)
+    owned = Qwen3_5RMSNorm(hidden, eps=1e-6)
+    _seed_weights(upstream, owned)
+
+    torch.manual_seed(0)
+    x = torch.randn(8, hidden, dtype=torch.bfloat16)
+    y_up = upstream(x.clone())
+    y_own = owned(x.clone())
+
+    torch.testing.assert_close(y_own, y_up, rtol=0, atol=0)
+    print("  [PASS] test_rms_block_level_equivalence")
+
+
+def test_rms_head_dim_equivalence() -> None:
+    """Head-dim (hidden_size=256) equivalence — the q_norm / k_norm shape."""
+    hidden = 256
+    upstream = GemmaRMSNorm(hidden, eps=1e-6)
+    owned = Qwen3_5RMSNorm(hidden, eps=1e-6)
+    _seed_weights(upstream, owned)
+
+    torch.manual_seed(1)
+    x = torch.randn(8, 24, hidden, dtype=torch.bfloat16)
+    y_up = upstream(x.clone())
+    y_own = owned(x.clone())
+
+    torch.testing.assert_close(y_own, y_up, rtol=0, atol=0)
+    print("  [PASS] test_rms_head_dim_equivalence")
+
+
+def test_rms_fused_residual_equivalence() -> None:
+    """Block-level fused-add-residual forward equivalence."""
+    hidden = 5120
+    upstream = GemmaRMSNorm(hidden, eps=1e-6)
+    owned = Qwen3_5RMSNorm(hidden, eps=1e-6)
+    _seed_weights(upstream, owned)
+
+    torch.manual_seed(2)
+    x = torch.randn(8, hidden, dtype=torch.bfloat16)
+    r = torch.randn(8, hidden, dtype=torch.bfloat16)
+
+    y_up, r_up = upstream(x.clone(), r.clone())
+    y_own, r_own = owned(x.clone(), r.clone())
+
+    torch.testing.assert_close(y_own, y_up, rtol=0, atol=0)
+    torch.testing.assert_close(r_own, r_up, rtol=0, atol=0)
+    print("  [PASS] test_rms_fused_residual_equivalence")
+
+
+def test_customop_registry_no_collision() -> None:
+    """Both 'qwen3_5_rms_norm' and 'gemma_rms_norm' must coexist."""
+    assert "qwen3_5_rms_norm" in op_registry, "owned CustomOp name missing"
+    assert "gemma_rms_norm" in op_registry, "upstream CustomOp name was clobbered"
+    assert op_registry["qwen3_5_rms_norm"] is Qwen3_5RMSNorm
+    assert op_registry["gemma_rms_norm"] is GemmaRMSNorm
+    print("  [PASS] test_customop_registry_no_collision")
+
+
+def test_mlp_class_surface() -> None:
+    """Qwen3_5MLP constructor surface matches spec (no expert_gate / reduce_results)."""
+    import inspect
+
+    params = list(inspect.signature(Qwen3_5MLP.__init__).parameters)
+    assert "expert_gate" not in params, "expert_gate should have been dropped"
+    assert "reduce_results" not in params, "reduce_results should have been dropped"
+    # Expected surface:
+    assert params == [
+        "self",
+        "hidden_size",
+        "intermediate_size",
+        "hidden_act",
+        "quant_config",
+        "prefix",
+    ], f"unexpected Qwen3_5MLP signature: {params}"
+    print("  [PASS] test_mlp_class_surface")
+
+
+def main() -> int:
+    print("Running Tier-1 smoke tests for nvllm Phase C owned layers:")
+    tests = [
+        test_rms_block_level_equivalence,
+        test_rms_head_dim_equivalence,
+        test_rms_fused_residual_equivalence,
+        test_customop_registry_no_collision,
+        test_mlp_class_surface,
+    ]
+    # CustomOp.__init__ calls get_current_vllm_config(); wrap the whole run
+    # in a bare VllmConfig context so module-import-time instantiations work
+    # without standing up a full engine. Same pattern as tests/conftest.py
+    # `default_vllm_config` fixture.
+    failed = 0
+    with set_current_vllm_config(VllmConfig()):
+        for fn in tests:
+            try:
+                fn()
+            except Exception as exc:  # noqa: BLE001
+                print(f"  [FAIL] {fn.__name__}: {exc}")
+                failed += 1
+    total = len(tests)
+    print(f"\nResult: {total - failed}/{total} passed.")
+    return 0 if failed == 0 else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tools/pre_commit/mypy.py

@@ -36,6 +36,7 @@
 EXCLUDE = [
     "vllm/model_executor/models",
     "vllm/nvllm/models",
+    "vllm/nvllm/layers",
     "vllm/model_executor/layers/fla/ops",
     # Ignore triton kernels in ops.
     "vllm/v1/attention/ops",

vllm/nvllm/layers/__init__.py

@@ -0,0 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the nvllm fork
+"""nvllm layer primitives — fork-owned RMSNorm, MLP, etc.
+
+Classes here are full copies of upstream class bodies (not subclasses).
+Upstream renames must not silently break fusion wiring. Only layer primitives
+the uber-kernel will fuse against live here — embedding, MoE-block, parallel-
+linear, and activation primitives remain upstream (see Phase C spec §Non-goals).
+"""
+
+from vllm.nvllm.layers.layernorm import Qwen3_5RMSNorm
+from vllm.nvllm.layers.mlp import Qwen3_5MLP
+
+__all__ = ["Qwen3_5RMSNorm", "Qwen3_5MLP"]