Escha-W2 port: 2-bit trellis decoded in the GEMV — Qwen3.6-35B-A3B and Qwen3.8-27B - #694

Open

nwoolmer wants to merge 101 commits into

Commits on Sep 2, 2026

docs: Escha-W2 port design

nwoolmer
and
claude
committed
79f5788
docs: verify Escha-W2 format claims against golden vectors

nwoolmer
and
claude
committed
4584359
docs: third review pass — coverage gap, unrunnable gate, sizing

nwoolmer
and
claude
committed
78d0717
docs: fourth review pass — Escha's own loader contract

nwoolmer
and
claude
committed
eef9e09
docs: fifth review pass — ran the MoE fixture

nwoolmer
and
claude
committed
276f4cc
docs: correct the quant-type id survey

nwoolmer
and
claude
committed
8f96230
docs: Escha-W2 Phase 1 implementation plan (35B-A3B)

nwoolmer
and
claude
committed
8180668
fix(plan): the G1 round-trip verifier did not work

nwoolmer
and
claude
committed
5e67627
plan: resolve three pre-flight rubric conflicts

nwoolmer
and
claude
committed
be9824a
feat(escha): trellis codec CPU reference, gated on golden vectors

nwoolmer
and
claude
committed
1ed615a
plan: escha f16 rounding is RNE, not hipfire's truncating helper

nwoolmer
and
claude
committed
8fd733b
feat(escha): H128 transforms and scale folding in the CPU reference
nwoolmer
committed
41106f3
test(escha): cover input/output transform ordering and fold_scales mixed cases
nwoolmer
committed
464b4f8
plan: Task 3 builds no moe_block — title and rationale corrected

nwoolmer
and
claude
committed
b971a41
feat(escha): expert linear, SwiGLU and w8a16 in the CPU reference
nwoolmer
committed
b75203a
fix(escha): expert_linear must not skip zero activations

nwoolmer
and
claude
committed
4fb616e
plan: drop expert_linear's zero-activation skip

nwoolmer
and
claude
committed
5f7b1a7
feat(escha): register ESCHA2T16/ESCHA3T16 and the EschaH128 rotation plan
nwoolmer
committed
765f46c
fix(escha): fail closed on Unimplemented arch predicate, explicit EschaH128 gemv_steps arm
nwoolmer
committed
fdd7ad7
fix(escha): gemv_steps returns Result instead of panicking on EschaH128
nwoolmer
committed
b29ceb2
feat(escha): safetensors -> hfq converter with the leaf contract enforced

nwoolmer
and
claude
committed
12856d4
plan: converter must embed the parsed config in .hfq metadata

nwoolmer
and
claude
committed
352760b
fix(escha): embed config in .hfq metadata, fail closed not panic, add convert_escha test

nwoolmer
and
claude
committed
fb82481

Commits on Sep 3, 2026

plan: Tasks 7-8 host wrappers used a launch API that does not exist

nwoolmer
and
claude
committed
213fa2c
feat(escha): G1 verbatim round-trip verifier + config load smoke test

nwoolmer
and
claude
committed
79ff405
feat(escha): GPU tile decode kernel, bit-exact against the CPU reference
nwoolmer
committed
fe2f128
fix(escha): correct kernel doc, validate decode inputs, drop select-cascade indexing, add device-resident path

nwoolmer
and
claude
committed
2d99ce6
feat(escha): H128 input/output transform kernels, gated against the reference
nwoolmer
committed
2a73edd
perf(escha): parallelise the H128 butterfly across all 128 threads
nwoolmer
committed
61e3ab8
docs(escha): narrow the __float2half sign-of-zero comment, widen the H128 bench

nwoolmer
and
claude
committed
76eb452
plan: Task 10 must batch H128 across experts (measured hard requirement)

nwoolmer
and
claude
committed
eae4ff1
test(escha): G4b — arch-6 router selects the same experts as escha
nwoolmer
committed
264933b
fix(escha): round router logits to f16 before top-k (escha only) + fix test predicate drift

nwoolmer
and
claude
committed
126a461
feat(escha): load escha experts as Q8_0 and run the H128 pair in arch-6

nwoolmer
and
claude
committed
ea9e9f9
fix(escha): resolve escha tensor names through the loader's candidate list

nwoolmer
and
claude
committed
f488b96
fix(escha): fail closed on the indexed GPU-top-K path, gate prefill, widen the knife-edge arm-1 bound

nwoolmer
and
claude
committed
ed0d21e
perf(escha): run the routed experts on the indexed GPU-top-K path

nwoolmer
and
claude
committed
5d552fb
fix(escha): embed the tokenizer, chat template and generation_config

nwoolmer
and
claude
committed
f1ecea5
feat(escha): registry entry, coherence and the G5 KLD result for Phase 1

nwoolmer
and
claude
committed
5e092e6
perf(escha): pack the routed experts into one buffer per (layer, projection)

nwoolmer
and
claude
committed
6547c78
perf(escha): stop computing logits for prefill tokens that discard them

nwoolmer
and
claude
committed
2e549c7
fix(rdna): stop the FP16 activation scratch returning another tensor's conversion

nwoolmer
and
claude
committed
d4dff4a
perf(escha): make batched prefill reachable — 45.3 -> ~163 tok/s

nwoolmer
and
claude
committed
06ab4db
fix(escha): make the G1 verifier fail when it compares nothing

nwoolmer
and
claude
committed
a63d04d
fix(rdna): delete HIPFIRE_FP16_X_LEGACY_CACHE

nwoolmer
and
claude
committed
4ef29bc
fix(dispatch): fail closed when an escha layer takes a non-escha prefill path

nwoolmer
and
claude
committed
571d3a9
fix(dispatch): dtype_needs_rotation must answer true for the escha dtypes

nwoolmer
and
claude
committed
2de41e4
fix(qwen35): refuse an escha MoE in the MTP head instead of skipping its transforms

nwoolmer
and
claude
committed
2dec2ff
perf/cleanup(escha): OnceLock the debug hooks, drop a divergent-barrier guard

nwoolmer
and
claude
committed
a5e37d1
test(escha): make the gates enforce their own contracts

nwoolmer
and
claude
committed
e113254
docs(escha): correct §10.3, record the gates and the limitations, register G1-G6

nwoolmer
and
claude
committed
ce22593
feat(escha): fused routed GEMV that decodes the trellis code in-register

nwoolmer
and
claude
committed
1d7d1ea
perf(escha): default the routed experts to the native trellis store

nwoolmer
and
claude
committed
02c36d3
perf(escha): expert-grouped GEMM for batched prefill

nwoolmer
and
claude
committed
e193b4d
fix(escha): grouping threshold must not depend on tile height

nwoolmer
and
claude
committed
7a85963
fix(qwen35): detect Escha-W2 before the EP-shard block

nwoolmer
and
claude
committed
2f8be99
perf(escha): WMMA grouped GEMM — measured, and it is NOT the prefill fix

nwoolmer
and
claude
committed
4cee172
perf(escha): WMMA grouped GEMM, with B read from global — prefill 375 -> 592

nwoolmer
and
claude
committed
1945a9e
perf(escha): measure WMMA at decode — negative, and it says why

nwoolmer
and
claude
committed
5642adc
feat(escha): dense down-quant tooling + the KLD, and a dispatch bug it found

nwoolmer
and
claude
committed
456748a
perf(escha): down-quant dense weights to MQ6, +17% decode
nwoolmer
committed
db226d9
fix(kernels): 1-byte overread past every 6-bit group, 9 WMMA kernels
nwoolmer
committed
7ebcdf7
refactor(escha): generalise HIPFIRE_ESCHA_DENSE parsing to <fmt>[:|!]<list>

nwoolmer
and
claude
committed
2b5a4c8
fix(kernels): gemv_mq6g256v2_multirow_r2 produced garbage off gfx1100/1201
nwoolmer
committed
a952a60
perf(escha): route gate_up through the wide expert GEMV, -8.7% on that kernel
nwoolmer
committed
5f7b676
feat(escha): three shipped dense recipes, q8 the default
nwoolmer
committed
3b6d92b
registry: three escha SKUs, local-only until the HF upload lands
nwoolmer
committed
6c26e8f

Commits on Sep 4, 2026

registry: publish the three escha SKUs
nwoolmer
committed
0e5cc10
registry: extension names the quant format, not the dense recipe
nwoolmer
committed
211d43e
fix(escha): converter silently dropped the 27B's MTP head
nwoolmer
committed
bc41962
feat(escha): dense-projection leaf naming and detection
nwoolmer
committed
28df9f2
feat(escha): load one dense escha linear (weight + rin/rout + bias)
nwoolmer
committed
d8a01b5
feat(escha): preflight accepts escha-coded dense projections

nwoolmer
and
claude
committed
0e20917
feat(escha): dense linear forward primitive (H128 in / GEMV / H128 out / bias)
nwoolmer
committed
1085ef9
fix(escha): dense forward must use the BATCHED H128, and a gate that proves it
nwoolmer
committed
40d6d06
test(escha): prove the H128 rotations fold into the weight
nwoolmer
committed
91d164f
feat(escha): fold H128 rotations and diagonals into the weight
nwoolmer
committed
6fe050b
feat(escha): converter fold mode — the 27B now runs end to end
nwoolmer
committed
ce3b4f3
feat(escha): load the 27B's per-projection biases
nwoolmer
committed
eacf76e
feat(escha): apply the 27B's biases — PPL 13.9142 -> 13.6957
nwoolmer
committed
a5f622e
feat(escha): biases on the full-attention prefill paths
nwoolmer
committed
7dd1712
fix(escha): H128/SwiGLU raced the capture stream — G5 is usable again
nwoolmer
committed
0f0776e
feat(escha): native trellis GEMV for dense linears — no decode, no fold
nwoolmer
committed
58df40b
test(escha): time the dense stores — native is 1.7x faster than Q8_0
nwoolmer
committed
64bd257
feat(escha): loader accepts trellis quant types 42/43
nwoolmer
committed
94f6f69
feat(escha): EschaProj — batched trellis projection, gate-verified
nwoolmer
committed
bcd3796
feat(escha): layers carry trellis metadata; loading side complete
nwoolmer
committed
210eeaf
feat(escha): native trellis 27B RUNS — decode path, gate passed
nwoolmer
committed
f554698
feat(escha): route native trellis through batched prefill
nwoolmer
committed
56aa9c8
perf(escha): grouped GEMM for batched prefill — 14.0 -> 71.8 tok/s
nwoolmer
committed
20707f7
bench: escha native GEMV at dense-decode shapes

nwoolmer
and
claude
committed
8a735af
fix(saddle-lab): stop the n-gram blocker banning the prompt's own tokens

nwoolmer
and
claude
committed
5863fd2
Merge remote-tracking branch 'warpfront/master' into nw_escha_w2
nwoolmer
committed
8fd307e
perf(escha): batched prefill was broken, then unamortised — 52 -> 108 tok/s

nwoolmer
and
claude
committed
d4d4e16
feat(infer): use batched prefill for text-only prompts

nwoolmer
and
claude
committed
9b77400
perf(escha): nt-major tile grid for dense codes, 24% on the decode GEMV

nwoolmer
and
claude
committed
bb77ff8

Commits on Sep 5, 2026

registry: escha SKUs on the xt/base/pro convention, and add the 27B

nwoolmer
and
claude
committed
7813a8a
registry: measured min_vram for the 27B escha SKUs

nwoolmer
and
claude
committed
388ffc2
registry: put both escha models' min_vram on the same measured basis

nwoolmer
and
claude
committed
df45930
feat(qwen35): load MTP heads carried as in-trunk mtp.* tensors
nwoolmer
committed
e1d2d55
perf(generate): add HIPFIRE_DECODE_PROFILE to the AR decode loop
nwoolmer
committed
33ba046