npu/forward: drop the unused fused-GU host path and kernel-internal comments

[bong-water-water-bong]

Follow-up to #179/#181/#209. The NPU host code on main carried the private NPU kernels' internals in public: the fused int4 GU kernel's tile layout, which byte regions the object-FIFO delivers, the on-chip dequant stage, the SRAM budget, REG_M tiling limits, generator versions and private issue numbers. Owner rule: NPU kernels, their designs and detailed write-ups stay private.

Removed (dead code on the public route; no callers outside npu_engine_i8ctx_inc.h):

• npu/forward/engine-src/gu_i4_pack.h and q4nx_raw.h (the fused int4 GU packer and its raw reader)
• the fused GU→SiLU→D block in npu_engine_i8ctx_inc.h: launch_fused, dequant_fused, fused_elf_capable, packB_into_fused*, update_fused_header*, fused_gs_off and its constants
• the unused bf16_pair field

Reworded: the remaining comments that named kernel internals now state only the host-side contract (M=128 streams, one instruction BO per context, zero-padded rows).

13 insertions, 1,027 deletions. No behaviour change: q4nx_forward.cpp, the only file including the header, compiles cleanly against XRT's headers on strixhalo (g++ -std=c++20 -fsyntax-only). The private fused path stays in npu-kernels. Git history still holds the removed text.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 8a1ec60

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 8a1ec60)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 179 - Partially compliant Compliant requirements: GGUF-on-NPU serving route implemented Model repacking to Q4NX format Support for Qwen2.5-7B, MiniCPM4-8B, MiniCPM5-1B Verification on real NPU with correct token outputs Fixes for MiniCPM4-8B including embedding_scale, residual_scale, logit_scale, LongRoPE Repack guard implemented to fail on critical scale-like scalars Cache-hit validation added logit_scale fidelity confirmed Non-compliant requirements: GLM-4.7-Flash (deepseek2, MLA + MoE) support not claimed MiniCPM4-8B tokenizer fallback to bytes due to BPE-only restriction Foreign Q4NX directory trust not addressed Provenance stamp not implemented Requires further human verification: Verification of GLM-4.7-Flash support Verification of MiniCPM4-8B tokenizer behavior Verification of foreign Q4NX directory trust behavior 181 - Partially compliant Compliant requirements: Prefer canonical npu-infer converter Forward re-sync from canonical npu-infer forward Fix streaming path's dangling capture Non-compliant requirements: None Requires further human verification: None 209 - Partially compliant Compliant requirements: Support for all six architectures GLM-4.7-Flash fixes Qwen3 MoE and qwen35moe support NPU serve takes /tmp/1bit-npu-device.lock Repack records GGUF arch for provenance stamp Tokenizer accepts Unigram tokenizer.json Registry rows for six GGUF-on-NPU arches Fix leading space stripping in streaming decode Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Dead Code Removal The PR removes dead code related to fused int4 GU kernel functionality, including gu_i4_pack.h, q4nx_raw.h, and associated functions like launch_fused, dequant_fused, fused_elf_capable, etc. This removal is correct as these are internal kernel details not needed in the public interface. return true; } // ── M is ALWAYS 128 — there is no valid M<128 instruction stream ── // The kernels are built for M=128, so gemm_generate_sequence_i8 ignores // its M argument: the stream depends only on (K, N). A stream built for a // smaller M does not complete. // // Smaller batches MUST reuse the M=128 stream and pass the real row count // as `am` to go()/launch_*: quantize_async zero-pads rows [am, 128) so only // rows [0, am) are valid. That is what npu_engine_universal, zaya_decode // and zaya_npu_runner do for single-token decode. Real small-M streams // require per-shape small-M xclbins (build_xclbins.sh Peano path), not a // runtime regen. (A former regen_insts(int M) re-uploaded an identical // M=128 stream and claimed to resize the batch — removed as misleading.) #else // Stub: npu_instr_utils.hpp not available — use init() with pre-gen'd files bool init_with_generator(xrt::device&, const char*, int, int, int, int) { fprintf(stderr, " I8Ctx: init_with_generator unavailable (no npu_instr_utils)\n"); return false; } #endif // ── Init: consume a FULL ELF directly (no xclbin) ── // Kernel signature: (bo0, bo1, bo2) == (A, B, C). No instruction BO: the // TXN control stream lives in the ELF's .ctrltext. Kernel name is // "<kernel>:<instance>" from the aiecc full-ELF config ("main:seq" for the // v27 designs). bool init_elf(xrt::device& d, const char* ep, int M, int K, int N, int nlayers, const char* kname = "main:seq") { MD = M; KD = K; ND = N; NL = nlayers; fprintf(stderr, " I8Ctx::init_elf ep=%s M=%d K=%d N=%d nl=%d\n", Comment Updates Comments that previously referenced internal kernel details have been updated to reflect only the host-side contract, such as M=128 streams and zero-padded rows. This improves clarity and prevents confusion about internal implementation details. int bC_nd = 0; // bo2 (C2) size override: MD*bC_nd*4 bytes when > 0, for a // kernel whose output is wider than the logical ND. std::unique_ptr<xrt::xclbin> xc;

[github-actions]

Persistent review updated to latest commit 8a1ec60