docs: add README build mode guidance

[zcxGGmu]

Summary

• Add README build-mode guidance for the inference quick start.
• Show the recommended prebuilt install, a CPU-only / portable AVX2 source build, and a CUDA source build path.
• Clarify that pure CPU experiments should use KT-Kernel CPU backends directly, while balance_serve remains a CPU-GPU heterogeneous serving path.

Refs #1187

Test Plan

• python README guard check for the new build-mode text
• ./install.sh --help output contains --manual
• kt-kernel/install.sh --help output contains CPUINFER_CPU_INSTRUCT and CPUINFER_ENABLE_AMX
• git diff --check HEAD~1..HEAD

[greptile-apps]

Greptile Summary

This PR adds a new "Installation / build modes" section to the root README, documenting three build paths: a prebuilt pip install, a CPU-only AVX2 source build, and a CUDA source build for heterogeneous serving. It also clarifies the scope of balance_serve (CPU-GPU path, not standalone CPU).

• CPU-only source build correctly goes into kt-kernel/ and calls ./install.sh all --manual, which runs install_dependencies before the build step — the previously flagged build --manual issue has been resolved.
• CUDA source build uses the root ./install.sh all, which calls install_deps (system dependencies via kt-kernel/install.sh deps) before delegating the build to kt-kernel/install.sh build, so system prerequisites are correctly installed on a fresh environment.
• Prebuilt path (pip install kt-kernel sglang-kt) is a straightforward one-liner that needs no further qualification.

Confidence Score: 5/5

Documentation-only change; all three build paths have been verified against the actual install.sh scripts and are correct.

The CPU-only build correctly uses all --manual (addressing the previously flagged issue), and the CUDA build's root ./install.sh all properly calls install_deps before delegating the build, so system prerequisites are installed on a fresh clone in both paths.

Files Needing Attention: No files require special attention.

Important Files Changed

Filename	Overview
README.md	Adds an 'Installation / build modes' section with three build paths (prebuilt, CPU-only source, CUDA source); commands are correct against the actual install.sh scripts

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User wants to install ktransformers] --> B{Build path?}
    B --> C[Prebuilt]
    B --> D[CPU-only source]
    B --> E[CUDA source]

    C --> C1["pip install kt-kernel sglang-kt"]

    D --> D1["git clone --recursive"]
    D1 --> D2["cd ktransformers/kt-kernel"]
    D2 --> D3["CPUINFER_USE_CUDA=0 CPUINFER_CPU_INSTRUCT=AVX2\n./install.sh all --manual"]
    D3 --> D4["install_dependencies (libhwloc-dev, pkg-config)"]
    D4 --> D5["build_step --manual (AVX2, no CUDA)"]

    E --> E1["cd /path/to/ktransformers"]
    E1 --> E2["CPUINFER_USE_CUDA=1 ./install.sh all"]
    E2 --> E3["init_submodules"]
    E3 --> E4["install_deps - kt-kernel/install.sh deps"]
    E4 --> E5["install_sglang"]
    E5 --> E6["install_kt_kernel - kt-kernel/install.sh build (CUDA)"]

Loading

_{Reviews (2): Last reviewed commit: "[docs]: use all for fresh CPU build mode" | Re-trigger Greptile}

[zcxGGmu]

Thanks for the review — I pushed 3c17240 addressing the fresh-clone build command feedback.

Changes made:

• Updated the CPU-only source build example to use ./install.sh all --manual so dependency installation runs before the build step.

Validation:

• grep -nF 'CPUINFER_USE_CUDA=0 CPUINFER_CPU_INSTRUCT=AVX2 CPUINFER_ENABLE_AMX=OFF ./install.sh all --manual' README.md — passed.
• git diff --check — passed.