refactor: slim PLENA compiler frontend and codegen#38

Merged

booth-algo merged 32 commits into

mainfrom

feat/codegen-addr-reg-init

May 13, 2026

Commits on May 11, 2026

feat(aten): automatic HF model → PLENA ISA compiler
booth-algo
committed
rename: model_compiler.py → plena_parser.py
booth-algo
committed
rename: plena_parser.py → plena_frontend.py
booth-algo
committed
feat(plena_frontend): native-dimension compilation with multi-head attention
booth-algo
committed
fix: audit fixes — assertions, docstring, VRAM padding
booth-algo
committed
feat(plena_frontend): on-chip K/V projections + RoPE
booth-algo
committed
feat(plena_frontend): true e2e — real embeddings + lm_head + multi-layer
booth-algo
committed
feat(plena_frontend): three-way comparison — HF float32 vs golden vs emulator
booth-algo
committed
feat: causal masking + real embeddings + lm_head + three-way comparison
booth-algo
committed
feat: add layer progress markers to ISA output for tracking
booth-algo
committed
fix: add load_toml_config to compiler/utils/load_config.py
booth-algo
committed
feat: golden_precision ablation for quantization gap proof
booth-algo
committed
fix(golden): BF16 truncation at all pipeline stages + K-split temp fix
booth-algo
committed
fix(golden): use MXFP8-quantized X as golden input, matching HBM
booth-algo
committed
feat: VRAM-in-the-loop stage comparison for emulator validation
booth-algo
committed
fix(ci): guard toml import for CI environments without toml package
booth-algo
committed
style: clean up informal comments in vram_stage_compare
booth-algo
committed
fix: use 'compiler' not 'PLENA_Compiler' for doc paths (post-merge fix)
booth-algo
committed
fix: revert to PLENA_Compiler path (matches production layout)
booth-algo
committed
refactor: split PLENA compiler DSL and ISA layers
booth-algo
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

refactor: slim PLENA compiler frontend and codegen#38

refactor: slim PLENA compiler frontend and codegen#38
booth-algo merged 32 commits into
mainfrom
feat/codegen-addr-reg-init

Commits on May 11, 2026

feat(aten): automatic HF model → PLENA ISA compiler

rename: model_compiler.py → plena_parser.py

rename: plena_parser.py → plena_frontend.py

feat(plena_frontend): native-dimension compilation with multi-head attention

fix: audit fixes — assertions, docstring, VRAM padding

feat(plena_frontend): on-chip K/V projections + RoPE

feat(plena_frontend): true e2e — real embeddings + lm_head + multi-layer

feat(plena_frontend): three-way comparison — HF float32 vs golden vs emulator

feat: causal masking + real embeddings + lm_head + three-way comparison

feat: add layer progress markers to ISA output for tracking

fix: add load_toml_config to compiler/utils/load_config.py

feat: golden_precision ablation for quantization gap proof

fix(golden): BF16 truncation at all pipeline stages + K-split temp fix

fix(golden): use MXFP8-quantized X as golden input, matching HBM

feat: VRAM-in-the-loop stage comparison for emulator validation

fix(ci): guard toml import for CI environments without toml package

style: clean up informal comments in vram_stage_compare

fix: use 'compiler' not 'PLENA_Compiler' for doc paths (post-merge fix)

fix: revert to PLENA_Compiler path (matches production layout)

refactor: split PLENA compiler DSL and ISA layers

Commits on May 12, 2026

refactor: slim ATen compiler path

refactor: split ATen frontend references

refactor: call PLENA compiler directly from ATen frontend

refactor: route HF frontend through aten ops

refactor: split memory state from matrix ISA

refactor: remove tile compiler compatibility alias

refactor: remove plena compiler compatibility facade

docs: add aten package tree

refactor: remove remaining compatibility shims

docs: refresh generator path comment

docs: refresh sim env path comments

refactor: move ATen e2e runner out of generator