[jschmied]

Closes #1666 once its base lands. Draft: it carries the commits of the PRs below it until they merge, so the diff shrinks to the last seven commits (from c682749c0 on).

Stacked on:

• xpu: ask Level Zero for the driver and device counts before the lists #1672 (device enumeration);
• xpu: Gen12LP (Iris Xe, DG1) builds Nemotron's kernels without DPAS #1673 and xpu: a Gen12LP server builds without the Qwen3.8-27B runner, and nem_prow gets an 8 x 4 image (stacked on #1673) #1692 (Gen12LP without DPAS);
• xpu: without Sysman (WSL2), free memory falls back to the core total (stacked on #1672) #1691 (free memory without Sysman, WSL2);
• lfm3: admit official MLX text and VL checkpoints #1513 (LFM3 admission), plus the module-name table suggested there (ecdc0f8cd);
• lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634's ref64.py, numpy path (a66a2bf75).

What it adds

• Kernels without DPAS: the short conv, QK norm with partial RoPE, gated GQA attention over a ring, the sandwich norm and SwiGLU, each with an FP64 op test and a planted control.
• The LFM3 forward on the Intel GPU:
  • prefill in windows, decode, and the conv state and KV rings;
  • weights loaded through names.zig, MLX 4-bit group 64 as stored;
  • a checkpoint test against ref64.py's float64 dumps.
• Serving: xpu-native serves LFM3 through the lane host. --parallel streams run on their own sequences, and a prompt cache keeps states.
• Exact speed work:
  • decode fuses w1 + w3 + SwiGLU, the four attention matvecs, and RoPE with the KV store;
  • prompts run on nem_prow_gen12 in windows of 128 rows.
• Docs: the Iris Xe row in docs/recipes/xpu-arc.md.

Checks (Iris Xe, WSL2)

• Forward against float64: every layer of 3 prompts is at 0.96–1.05× the bf16 gap. The 64 greedy tokens match the CUDA engine (lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634, MLX 4-bit) on all 3 prompts, and match again after every speed change, at windows of 128 and 64, with each fusion on and off.

• Streams: at --parallel 1, 2 and 4, 5 concurrent replies are byte-equal to solo, in two runs each.

• Prompt cache: a resumed second turn is byte-equal to the cold one, at 1k and 8k tokens. TTFT is 0.56 s against 13.25 s at 1k, and 0.80 s against 122.6 s at 8k.

• Speed, server, --context 16384:

  	before the speed work	after
  decode	14.8–15.8 tok/s	16.7–16.9 tok/s
  TTFT at 1k	13.1–14.2 s	11.7–12.2 s
  TTFT at 8k	130.8–131.3 s	116.3–117.7 s

• Spark (GB10): lean reports no new findings; zig build test, the contract, golden, security and fuzz steps pass; git diff --check is clean.

Overlaps

atc overlap reports file conflicts with:

• lfm3: add native Metal vision support above the text decoder #1502 (LFM3 vision on Metal, 6 files): both build on the LFM3 series;
• Gemma 4 E4B on the Intel Arc Pro B70 (Xe2) #1552 (Gemma 4 on Arc B70, 4 files): both touch xpu/.

I'll rebase onto whichever lands first.