[noahgift]

D3-SPLIT B of #4620: QK-norm models take the FP16 batched prefill when it fits (root 2)

Stacked on #4620 (A), and C (ec/4621-chat-session-kv, apr chat device KV) is stacked on this PR. The base branch is ec/3715-lambda-cells, so ci.yml does not run here. It runs once A merges and this PR retargets to main. The mutant cap is verified in that run.

What changes

• GpuProfile::disable_fp8_for_qk_norm still turns FP8 off for a QK-norm model (Qwen3: the FP8 batched prefill fails CPU parity, FP8 batched prefill fails CPU parity at the post-prompt decode step — catastrophic on per-head-QK-norm models (Qwen3-8B cos −0.10), marginal on Qwen2.5 (0.92) #3483). It no longer forces the serial prefill. Serial is one forward_gpu_resident per prompt token: Qwen3-1.7B could not prefill a 20k prompt within the 600 s cell timeout. The batched prefill is now chunked (fix(cuda): #3715 lambda run/chat cells: chunked prefill, chat KV, Qwen3.5 F16/IQ4_XS dequant, 8B fit #4620), and chunking is what bounded the old OOM.
• GpuProfile::serial_when_fp16_prefill_does_not_fit: a QK-norm model keeps the serial prefill when weights + FP16 cache + KV + prefill reserve exceed free VRAM. Qwen3-8B on 24 GB is the case that OOMed into a CPU fallback without this check. BATCHED_PREFILL set explicitly still overrides.
• OwnedQuantizedModelCuda::resident_estimate and the new gguf/cuda/resident.rs read the projection tensors themselves (bytes, in×out), not a dims formula, so the estimate cannot drift from the model.

Evidence

• Mutants on this diff (vs A, as CI counts them): 58 (gpu_profile 17, mod.rs 29, resident.rs 12). Cap is 60.
• Intel @451bc77c52 (--features cuda --lib): 17910 passed, 6 failed, 76 ignored. All 6 failures are device or ptxas tests (test_new_device_0, test_num_devices, test_cov002_cuda_likely_available, test_cuda_executor_new, gemm_bt_tiled_ptx_…, every_emitted_kernel_assembles), and intel has no NVIDIA driver (nvidia-smi fails), so they say nothing about this diff. fmt --check passes.
• clippy --features cuda -D warnings: 2 borrow_deref_ref errors at cuda/executor/residual.rs:314,343, already on main (from perf(qwen35-cuda): DP4A requant staleness (#4258) + argmax first-of-ties (#4259) + allocation-free decode plumbing (#4215) — replaces #4322 #4332 #4439), outside this diff. The receipt ec ran on the B files alone (resident:: pmat3477 fit_check gpu_profile) is 40/40.
• Device red/green for root 2: handoff/3715-fix-ec.md (ec, combined binary apr-fix4). The tensor-walk refactor of the resident estimate has not been re-measured on a device.

Refs #4621, Refs #3715. This PR does not close #4621: C closes it when it merges last.

Agent: aprender-e6 (split owner; code by aprender-ec)

🤖 Generated with Claude Code

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4621",
 "head": "451bc77c528822417a53b25190d7670dadd813eb",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 1
  },
  {
   "lane": 3,
   "verdict": "NO-VERDICT",
   "findings": 0
  }
 ]
}

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4621",
 "head": "451bc77c528822417a53b25190d7670dadd813eb",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 3
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 3
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

quorum-review (AD-04): three PASS — agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4621",
 "head": "1c97bea00e1f42564a6c1eeadb3a254c6ecd8e2f",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 1
  }
 ]
}

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4621",
 "head": "dcee3dd3ae9286ad88a1a7f4311c126000dccb4a",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 2,
   "verdict": "NO-VERDICT",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4621",
 "head": "dcee3dd3ae9286ad88a1a7f4311c126000dccb4a",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 1
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 2
  }
 ]
}

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4621",
 "head": "a6abbede70bb7b9d15013fd09e8c75d2341eef3b",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "FAIL",
   "findings": 3
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

Closed with a pointer to #4623 (0.70.2 backlog, aprender-78): this PR's #4621 fix (QK-norm models take the FP16 batched prefill, root 2) is carried by #4623 as cd324d0. Measured: crates/aprender-serve/src/cuda/executor/layers/prefill.rs and crates/aprender-serve/src/gguf/cuda/resident.rs are byte-identical on both heads. The three further commits on this branch (596b6bb, 6cac047, 17f499b) change the mutants gate. They belong to the mutants workstream, not to chat KV, so they are not folded into #4623. Branch ec/4621-fit-and-chat-kv is kept with them.