Four changes, each measured on a Xeon 8480C with Spark-X2.5-1.7B at 24
threads. Numbers are best-of-interleaved-passes; this host is shared and a
single sequential grid swung the llama.cpp control 30%, so anything under
~8% is not treated as a result.
**vLLM's CPU backend does have a 4-bit kernel.** An earlier note in this
tree said it did not, on the strength of `woq_int4_linear`,
`int4_scaled_mm_with_quant` and `da8w4_linear` all being absent. Those are
simply not the names it uses: `CPUWNA16LinearKernel` dispatches GPTQ/AWQ
checkpoints to `int4_scaled_mm_cpu`, an AMX W4A8 GEMM, whenever activations
are bf16 and the checkpoint carries no act-order. Four-bit weights times
eight-bit activations is the arithmetic llama.cpp's Q4_K_M already does.
`gptq_cpu_export` builds checkpoints for it; the kernel pins the zero point
at 8, so the grid must be symmetric, and a per-group scale search recovers
most of what that costs.
**A per-device weight path** (`edge/weight_path.py`). The fast kernel on each
device needs a different, mutually exclusive weight layout, so the choice has
to come from a probe rather than a flag: AMX x86 -> W4A8/GPTQ; x86 without
AMX, or priority="fidelity" -> W4A16/tinygemm (bf16 activations, the more
faithful of the two); ARM -> W4A16 but flagged unverified, because PyTorch's
4-bit CPU GEMM dispatches a different kernel there than the AVX-512 one
measured here; phone -> runs_in_vllm=False, the deliverable is an exported
graph with calibrated w8a16.
**The CPU compile cache never loaded.** `_try_load_aot_compiled_fn` finishes
with `compiled_fn.finalize_loading(cfg)`, which the CUDA wrapper object has
and the inductor CPU backend's plain function does not. The AttributeError
was swallowed as a cache miss, so every engine start recompiled the model:
warm start 49.8 s against 14.9 s with the missing no-op attached. Self-
disabling if upstream gives the CPU path a real one.
**The head gate stops being a GEMM.** A 2048x8 projection cannot amortize a
thread fan-out: dispatched as `aten::mm` it costs 1.155 ms of a decode step,
nearly all barrier. As an eager reduction behind an opaque op it costs 0.444
ms. It has to be opaque -- written inline the reduction is traced, and
inductor's loop made the whole step 1.8x slower (78.8 -> 44.5 tok/s).
Also included and *off by default*: `cpu_ir_ops` registers the fused
`_C::rms_norm` / `_C::fused_add_rms_norm` kernels that vLLM's CPU build ships
but gates on `GPGPU_DEVICE`. In isolation they are 3.6x and 4.5x faster than
the native path; end to end they are slightly slower (72.7 vs 74.0 tok/s),
because inductor was already fusing the native norm into its neighbours and
an opaque call breaks that fusion. Kept with the measurement recorded so the
next person does not repeat it. `gelu_and_mul` is deliberately not registered:
its fused CPU kernel is 2.7x slower at Spark's GEGLU shape.
Decode, W4A16 g64 -> W4A8 g128 with the gate fix: 69.1 -> 84.5 tok/s (+22%),
against llama.cpp Q4_K_M's 90.0; prefill 3347 against 1121 (3.0x). The two
4-bit paths are not equally faithful: against each build's own bf16 over 12
greedy prompts, Q4_K_M reproduces 4/12, W4A16 2/12 and W4A8 1/12.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Exports one Spark decode step with the KV cache as explicit tensors, sized
per layer type: the 21 sliding layers hold a fixed 512 entries at any
context, the 7 full-attention layers grow. That asymmetry is the whole
reason the architecture is interesting on a phone, so the two types get
separately sized caches, masks and rotary angles.
Verified against the HF reference at 1024 context: 138.5 dB logits SNR,
identical top-1 and top-5.
Profiled on a real Galaxy S25 (Snapdragon 8 Elite) through AI Hub, QNN
context binary on the NPU, ms per layer:
sliding (512 KV, any ctx) fp16 1.689 w8a16 0.912 w4a16 0.917
full @ 1k fp16 1.559 w8a16 1.129 w4a16 1.127
full @ 4k fp16 2.964 w8a16 1.681 w4a16 1.662
Composed into a step (21 sliding + 7 full), w4a16 gives 27.2 ms at 1k and
30.9 ms at 4k, against 31.6 and 46.5 ms for the same model with every layer
full-attention -- 1.16x and 1.51x. Worth noting the hybrid is 0.94x at 1k in
fp16: sliding layers rotate all 256 head dims where full layers rotate 64,
and below roughly 2k that RoPE cost outweighs the shorter cache.
Also moves the head-wise attention gate off vLLM's CPU linear dispatch. The
gate is [2048, 8], and any linear that narrow is routed to the sgl-kernel
packed GEMM, whose fake-tensor rule assumes the packed weight keeps N in dim
0; VNNI packing does not at that size, so it reports K and torch.compile
fails tracing the gate multiply. A plain parameter with F.linear gains
nothing less -- a 2048x8 matmul has no use for packed AMX tiles -- and it
keeps the model compilable on CPU and GPU alike.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
Spark-X2.5-1.7B is an Apache-2.0 on-device model built for agentic tool
use. vLLM 0.28 has no implementation for it, so it could not be served at
all; llama.cpp already ships one, which made it the only edge engine able
to run the model.
Three traits separate it from a Llama-style decoder:
* Hybrid attention -- layer_types interleaves three 512-token sliding
layers with one full-attention layer, which is what makes the 1M-token
context affordable on a phone.
* Per-layer-type RoPE -- sliding layers rotate all 256 head dims with
theta 1e4, full layers only the first quarter with theta 5e6.
* A head-wise attention output gate -- g_proj reads the same post-norm
input as QKV and scales each head's attention output by a sigmoid.
The checkpoint also stores QKV pre-fused as a single q_k_v_proj, so
load_weights splits it before vLLM's sharded QKVParallelLinear sees it.
Verified against the official remote-code implementation on transformers
4.57.1 (the version config.json targets; it does not import on the 5.x in
our venvs). In fp32 all three scenarios match exactly -- 32/32 greedy
tokens with a max top-20 logprob difference of 2.2e-05. In bf16 the short
and the 2465-token long-context prompts still match 32/32, which is what
exercises the sliding layers; the tool-calling prompt diverges at token 23
on a near-tie, and fp32 confirms that is rounding rather than a defect.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
Noise is drawn eagerly with the request generators (same RNG consumption as the
eager loop) and passed in, so the compiled post-step is generator-free and
shape-static; used on the fp32/torchao CPU path outside stored (top-p) mode.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
In-situ parity against the bf16 model (VLLM_OMNI_PREDICTOR_PARITY_CHECK=1):
fp32 control keeps 93.8 % top-1 / 95 % top-5 / KL 0.0015; torch.ao dynamic
int8 falls to 22 % top-1 (max logit diff 7), so per-tensor activation
quantization is not acceptable as a default. Adds the ao-nohead variant
(int8 transformer body, fp32 lm heads) for measurement.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
torchao's dynamic-activation int8 is numerically wrong under torch.compile on
CPU (cosine 0.02 vs fp32; caused runaway generations), so it is not offered;
weight-only + compile is faithful but slower (78 ms vs 36 ms per frame for the
15 sub-steps), so torch.ao dynamic int8 stays the default. Parity mode now also
records top-5 overlap and KL vs the bf16 model; compiled-step equivalence test.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Measured on a Xeon 8480C (12+4 threads, 0.6B): the 15 residual-codebook sub-steps
re-ran the 5-layer predictor over the whole prefix each time (97 ms per frame of
the 134 ms talker step) and vLLM's stock sampler spent 9 ms per step in its
torch.compile'd random-sample kernel on CPU.
- code predictor: forward_cached (KV cache per layer, 2-token prefill then one
token per sub-step) and a shape-static forward_static_step (tensor position +
boolean mask) that torch.compile handles as one graph; incremental decoding
is the default on CPU (VLLM_OMNI_PREDICTOR_INCREMENTAL), optional compile
(VLLM_OMNI_PREDICTOR_COMPILE=step), optional thread cap.
- CPU fast path: dynamic int8 Linear layers (fp32 activations) for the predictor
and its lm heads (VLLM_OMNI_PREDICTOR_INT8, default on; micro-benchmark 70 ->
36 ms per frame for the 15 sub-steps); VLLM_OMNI_PREDICTOR_PARITY_CHECK=1
records top-1 agreement / max logit diff against the bf16 model in situ.
- Qwen3-TTS talker: prefer_model_sampler on CPU with a sample() that mirrors
vLLM's Sampler (logits processors + penalties, temperature, top-k/top-p,
exponential-noise sampling with per-request generators) without the compiled
kernel (VLLM_OMNI_CPU_FAST_SAMPLER, default on).
- step_stats: sampler phase timers and sample-metadata timers.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Counters model.forward_ms / model.sample_ms / model.talker_mtp_ms and
predictor.substep_ms / predictor.loop_ms (VLLM_OMNI_STEP_STATS_DIR); with
VLLM_OMNI_STEP_STATS_SYNC=1 the accelerator is synchronized around the model
timers so async kernels are attributed (attribution runs only).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
