models(transformer): wire Gemma 3 per-layer SW + attn/final softcaps + QPAS - #318

Draft

jamesburton wants to merge 73 commits into

kkokosa:mainfrom

jamesburton:issue/274-gemma3-transformer-wiring-rebased

Commits on Jun 6, 2026

core(position-encoding): ALiBi (Attention with Linear Biases) kernel infrastructure (#149)

jamesburton
and
claude
committed
8f5c19d
vulkan: scaffold cross-vendor GPU compute backend (#155)

jamesburton
and
claude
committed
112b0c3
loader(safetensors): binary parser foundation — File, DType, Descriptor (#154)

jamesburton
and
claude
committed
5425aa9
vulkan: baseline F32 + Q8_0 compute kernels (#167)

jamesburton
and
claude
committed
3f152c3
loader(safetensors): HfConfigExtractor + TransformerWeightsSafetensors for dense LLMs (#166)

authored and
jamesburton
committed
12f8475
lora(core): LoRA Phase 4a foundation — core types + PEFT loader + CPU forward path (#164)

jamesburton
and
claude
committed
9c55eb2
core(moe): Mixtral-family MoE foundation — top-k routing + per-expert SwiGLU (#175)

authored and
jamesburton
committed
5771e22
vulkan: add matmul_q8_0_gemm kernel (#173)
jamesburton
committed
7892168
vulkan: add bias_add_f32 kernel (#173)
jamesburton
committed
1df30a3
vulkan: fix matmul_q8_0 GEMV stride bug at K=32 with M>1 (#173)
jamesburton
committed
059126c
core(mla): MLA attention foundation — CPU PoC kernel + DeepSeek-V2/V3 config detection (#176)

jamesburton
and
claude
committed
7529125
core(mla): wire DeepSeek-V2/V3 forward path end-to-end (#176)

jamesburton
and
claude
committed
27913d0
core(mla): apply YaRN softmax mscale² correction (P2.2) (#178)

Your Name
authored and
jamesburton
committed
a7451f2
core(mla): Phase A expanded KV-cache (P2.3) (#178)

Your Name
authored and
jamesburton
committed
960e567
core(mla): Phase B latent KV-cache + absorbed attention (P2.3) (#178)

Your Name
authored and
jamesburton
committed
fc86766
kernels(cpu): vectorise MLA attention inner loops via TensorPrimitives (#178)

Your Name
authored and
jamesburton
committed
d8d5442
core(mla): Phase C hybrid cache — prefill expand + decode absorbed (P2.3) (#178)

Your Name
authored and
jamesburton
committed
718061f
core(moe): Qwen-MoE — shared-expert support for Qwen1.5 / 2 / 3-MoE (#180)

authored and
jamesburton
committed
7dc859a
Merge remote-tracking branch 'origin/issue/166-safetensors-hf-config-extractor-dense-loader' into issue/181-lora-parity-tinyllama-tests
jamesburton
committed
699c1e3
lora(tests): forward parity + multi-adapter switch tests (Phase 4a) (#181)

jamesburton
and
claude
committed
b160ae5
lora(tests): TinyLlama real-adapter integration test (Phase 4a) (#181)

jamesburton
and
claude
committed
75a53e3

Commits on Jun 7, 2026

lora(server): API integration — request field + admin endpoints (#189)

authored and
jamesburton
committed
794009f
lora(server): multi-adapter batcher + endpoint/batcher unit tests (#189)

authored and
jamesburton
committed
378102b
core(moe): DeepSeek-V2/V3 multi-shared-expert support (#194)

jamesburton
and
claude
committed
4610741
vulkan: scaffold transformer model, weights, KV cache, and forward state (#205)
jamesburton
committed
355e98d
vulkan: reset descriptor pool after each Launch (#205)

Your Name
authored and
jamesburton
committed
d4c4de1
vulkan(perf): device-local weights + KV cache via staging (#206)

authored and
jamesburton
committed
97e1366
vulkan(perf): fence-pipelined forward, one submit per forward (#206)

authored and
jamesburton
committed
3767a98
vulkan(perf): descriptor-set cache per kernel (#206)

authored and
jamesburton
committed
d184c17
vulkan: device-local activation scratch buffers (#206)

Your Name
authored and
jamesburton
committed
9162a5c
Merge #164 (LoRA-1 core types) into base for Vulkan LoRA Phase 4b
jamesburton
committed
01101a7
Merge #166 (safetensors HF dense loader) into base for Vulkan LoRA parity tests
jamesburton
committed
5687eef
vulkan(lora): VulkanLoraAdapter device upload + handle cache — Phase 4b (#216)

authored and
jamesburton
committed
68c3975
vulkan(lora): LoRA delta dispatch in Forward + per-projection wire-up — Phase 4b (#216)

authored and
jamesburton
committed
aac201a
vulkan(lora): real TinyLlama LoRA parity vs CPU — Phase 4b (#216)

authored and
jamesburton
committed
3c447cc
engine: IModel.ForwardBatch API + per-sequence-loop default impl (#222)
jamesburton
committed
b075793
docs(scheduling, server): continuous-batch scheduler details (#222)
jamesburton
committed
7c2401f
Merge remote-tracking branch 'origin/issue/216-vulkan-lora-adapter' into issue/229-vulkan-lora-fused-delta
jamesburton
committed
d37ec8e
vulkan(lora): fused delta GEMV shader (#229)

Your Name
authored and
jamesburton
committed
3a12104
vulkan(lora): kernel wrapper + tests (#229)

Your Name
authored and
jamesburton
committed
be00f3f
vulkan(lora): dispatch refactor — fused delta path (#229)

Your Name
authored and
jamesburton
committed
4aa7982
Merge origin/issue/178-mla-kv-cache-phases-mscale-simd into base for LoRA-5
jamesburton
committed
49c9dda
Merge origin/issue/194-moe-deepseek-multi-shared-expert into base for LoRA-5
jamesburton
committed
939c56b
LoRA: F16/BF16 native adapter weights + MLA/MoE acceptance + perf bench (Phase 4d) (#238)

jamesburton
and
claude
committed
7cebb66
LoRA: wire MLA and MoE deltas (#238)
jamesburton
committed
6c85a4b
Merge remote-tracking branch 'origin/issue/229-vulkan-lora-fused-delta' into issue/PLACEHOLDER-lora-phase-4d-q8-outerprod
jamesburton
committed
5474505
LoRA bench: synthetic adapter fixture (#243)

jamesburton
and
claude
committed
24c6697
LoRA bench: macro-benchmark BDN class (#243)

jamesburton
and
claude
committed
4258c86
LoRA Phase 4d.4: spike decision — Q8_0 LoRA delta path (B-only) (#243)

jamesburton
and
claude
committed
f7fb7e8
LoRA Phase 4d.4: Q8_0 LoRA-B kernel + adapter-load quantiser + parity tests (#243)

jamesburton
and
claude
committed
89d1033
LoRA Phase 4d.4: Q8_0 macro-bench wiring + asymmetric A/B dtype plumbing (#243)

jamesburton
and
claude
committed
8d8fcee
LoRA Phase 4d.4: Q8_0 path — rescue with dequant-once + macro-bench results (#243)

jamesburton
and
claude
committed
82ccfb0
LoRA Vulkan: Q8_0 upload path + dequant-on-load (Phase 4d.5 / Gap 1) (#243)
jamesburton
committed
b979806
LoRA CPU: pre-quantise-x once per layer (Phase 4d.5 / Gap 2) (#243)
jamesburton
committed
2f54c05
LoRA CPU: gate Q8_0 pre-quant-x fast path on env var (Phase 4d.5 follow-up) (#243)
jamesburton
committed
4a02d1f
LoRA: outer-product stage-2 fast path + lazy transposed-A cache (Phase 4d.6) (#243)
jamesburton
committed
5110e67
kernels(cpu)(attention): softCap parameter (Gemma 2/3 attn_logit_softcapping) (#255)

jamesburton
and
claude
committed
11d46a3

Commits on Jun 8, 2026

Commits on Jun 9, 2026

Commits on Jul 30, 2026

Merge remote-tracking branch 'refs/remotes/origin/issue/274-gemma3-transformer-wiring-rebased' into issue/274-gemma3-transformer-wiring-rebased
jamesburton
committed
c0ff1d0