feat(llama): add DecoderGgufMemSegConverter for the DSL inference path

[michalharakal]

Summary

• Adds a post-load converter targeting the DSL path's DecoderGgufWeights<T, V> format. This is the bind-time piece Phase 1B needs so quantized weights can flow through OptimizedLLMRuntime without dequantizing the entire model to FP32.
• Purely additive — 2 new files in :llm-inference:llama, no existing code modified. The legacy MemSegWeightConverter (operates on LlamaRuntimeWeights) stays in place and gets deleted in Phase 4 along with LlamaRuntime.

Why a separate converter

The original Phase 1B plan said "move MemSegWeightConverter to :llm-core (it's generic)". That premise was wrong: the existing converter operates on LlamaRuntimeWeights<FP32> (the hand-coded LlamaRuntime's named-field layer struct: wq, wk, wv, …) and is genuinely tied to that legacy format. The DSL path uses DecoderGgufWeights<T, V> — a flat tensor-name → tensor map — which has a different shape. So the DSL needs its own converter; the two coexist until Phase 4 deletes the legacy one.

Behavior

Input quant type	Output
Q4_0	Wrapped as Q4MemorySegmentTensorData
Q8_0	Wrapped as Q8MemorySegmentTensorData
Q4_K / Q5_K / Q6_K	Dequantized to FP32 (same trade-off MemSegWeightConverter makes; packed K-quant kernels aren't on the DSL hot path)
FP32 (no entry in quantTypes)	Pass-through unchanged
Anything else	Warning logged, pass-through (forward will fail at matmul if hit)

quantTypes is cleared on the result — packed tensors carry their own marker, dequantized tensors have no quant identity, and a stale map would mislead later consumers.

Why no pre-transpose (unlike the legacy converter)

The DSL's linearProject(input, weight) always calls ops.matmul(input, ops.transpose(weight)). For Q4/Q8 MemSeg tensors upstream's transpose is a shape-only metadata swap (free), so pre-transposing brings no benefit. For dequantized K-quants and FP32 tensors a runtime transpose still has a real cost — addressing it requires adding a pre-transposed marker that linearProject checks, which is tracked as a follow-up perf optimization in the sub-plan.

What ships next on top of this

• 1B.2 — fromGgufNative(...) entry point on QwenNetworkLoader / LlamaNetworkLoader / VoxtralNetworkLoader that loads with NATIVE_OPTIMIZED and runs the converter before binding into the DSL.
• 1B.4 — DSL-vs-LlamaRuntime numerical parity test on the same packed weights.
• After both: Phase 4 (CLI swap, delete LlamaRuntime).

Test plan

• 4 new tests in DecoderGgufMemSegConverterTest: empty-quantTypes no-op, Q4_0 wrap (verified via Q4MemorySegmentMarker), Q8_0 wrap, key-set preservation.
• No regressions: existing MemSegWeightConverterTest, all :llm-core, :llm-inference:llama, :llm-inference:qwen, :llm-runtime:kllama tests pass.
• CI green on PR.

Refs the Phase 1B sub-plan (~/.claude/plans/snazzy-wibbling-dewdrop-1B.md) and the closed #46.

🤖 Generated with Claude Code