e1 NPU quantization pipeline

The quantization pipeline produces calibration manifests consumed by the elizanpu IREE backend. Five formats target the NPU's hardware opcodes:

Format	Hardware path	Default use
PTQ INT8 (per-channel weights / per-tensor activations)	GEMM_S8, DOT4_S8	dense default for most CNN / small transformer
AWQ INT4 weight-only	DOT8_S4	LLM weights (best PPL at 3-4 bit)
GPTQ INT4 weight-only	DOT8_S4	fallback for non-LLM small-batch
FP8 E4M3	DOT4_FP8_E4M3 (scalar contract today; tensor path BLOCKED)	long-context LLM where INT8/INT4 PPL degrades
2:4 structured sparse INT4	SDOT4_S4_2_4	dense matmul layers with 50% magnitude pruning
INT2 BitNet	DOT16_S2 (scalar contract today; tensor path BLOCKED)	experimental ultra-low-precision LLM

Manifest schemas

Every calibrator emits a JSON manifest with a versioned schema string:

• eliza.ptq_int8_manifest.v1
• eliza.awq_int4_manifest.v1
• eliza.gptq_int4_manifest.v1
• eliza.fp8_e4m3_manifest.v1
• eliza.sparse_2_4_int4_manifest.v1
• eliza.int2_bitnet_manifest.v1

The IREE backend dispatches on the schema string at compile time.

Calibration flow

1. Run a representative batch through the model in fp32.
2. For each weight tensor, collect per-channel max-abs.
3. For each activation tensor, collect a sample of absolute values.
4. Call the matching calibrator's record_* methods.
5. Call build_manifest() and write the JSON to disk.
6. Pass to iree-compile --iree-input-quantization-manifest=<path>.

Status

• All six calibrators committed under compiler/quantization/.
• Unit tests pass (8/8) in repo CI without torch installed.
• Real PyTorch model integration: BLOCKED on torch + IREE inside the canonical Linux container.

Evidence gate

docs/evidence/compiler/quantization-evidence.yaml.