Qwen3.5-4b BF16 optimization and correctness (`sm_120`) - #266

Merged

localai-bot merged 17 commits into

mudler:mainfrom

richiejp:row/KERNEL-SSM-MAMBA

Aug 11, 2026

Commits on Aug 10, 2026

spec: define sm120 Qwen3.5 Pareto campaign
richiejp
committed
c66c24c
perf(cuda): restore sm120 GDN experiment arms (#206)
richiejp
committed
c7be732
test(cuda): guard post-conv token-tile fallback (#206)
richiejp
committed
41b0095
spec(bench): correct Qwen3.5 frontend timing gate (#206)
richiejp
committed
9f8dfd3
perf(bench): pretokenize prompts before TTFT timing (#206)
richiejp
committed
67767c2
test(bench): harden pretokenization contract (#206)
richiejp
committed
6b98621
spec(engine): require atomic benchmark wave admission (#206)
richiejp
committed
0473cd1
perf(bench): admit request waves atomically (#206)
richiejp
committed
b3f4b48
spec(cuda): bound argmax scratch growth (#206)
richiejp
committed
70f01f3
KERNEL-SAMPLING add opt-in geometric argmax scratch
richiejp
committed
8760336
fix(KERNEL-SAMPLING): bind argmax scratch adapter
richiejp
committed
89b930c
perf(KERNEL-SSM-MAMBA): add BF16 vector writeback discriminator
richiejp
committed
98878c6
fix(KERNEL-SSM-MAMBA): bind BF16 vector writeback contracts
richiejp
committed
ae85186
perf(KERNEL-SSM-MAMBA): close sm120 Pareto experiments
richiejp
committed
96919cc
spec(KERNEL-SSM-MAMBA): isolate BF16 cold-prefill TTFT
richiejp
committed
db3336f
perf(KERNEL-SSM-MAMBA): reject BF16 pre-serve warmup
richiejp
committed
fb0fc14
test(records): sync STATUS ratchet ceiling (#206)
richiejp
committed
a0a8a5b