feat(glm53): add device-resident HIP backend for MI350P

[azharlabs]

Summary

This PR adds a device-resident HIP GPU backend for GLM-5.3-Flash, optimized and validated on AMD Instinct MI350P.

The implementation keeps the main inference pipeline on the GPU, including embeddings, normalization, mHC, KDA, MLA/DSA attention, MoE routing, expert execution, final collapse, and the LM head.

Key changes

• Added GLM53_BACKEND=auto|gpu|cpu backend selection
• Added GPU capability detection and startup validation
• Implemented device-resident KDA with persistent recurrent state
• Implemented MLA/DSA with paged latent and indexer caches
• Added deterministic top-k expert routing
• Added GPU shared and routed expert execution
• Added a double-banked device expert cache with event-based publication
• Integrated coli_glm53_gpu_forward() into the production request path
• Added request-level failure handling and safe CPU fallback behavior
• Fixed grouped fmt=4 Int4 decoding using offset-binary nibble - 8
• Added component, integration, pipeline, and quality-gate tests

New files

• c/glm53_gpu.c
• c/glm53_gpu.h
• c/tools/glm53_quality.c
• GPU context and backend-selection tests
• mHC, KDA, MLA/DSA, and MoE GPU tests
• Full 45-layer pipeline tests
• Production request and fallback integration tests

Validation results

Validation using a real GLM-5.3 checkpoint on MI350P produced:

• Relative mean NLL delta: 6.65e-6
• Bitwise GPU determinism: Passed
• CPU/GPU generated-token agreement: Passed
• GPU pipeline startup and backend selection: Passed
• Int4 checkpoint decoding validation: Passed

Additional checks after rebasing onto the latest upstream dev:

• make -C c glm53: Passed
• Portable C test suite: Passed
• Python suite: 722 tests executed
• 38 tests skipped due to environment or optional dependencies
• 2 subprocess shutdown tests timed out in the restricted test environment:
  • test_stop_kills_only_the_requested_server
  • test_completion_stops_only_its_own_converter

The HIP test suite was not rerun in the PR preparation environment because MI350P hardware and hipcc were unavailable.

MI350P configuration

export HSA_OVERRIDE_GFX_VERSION=9.4.2
export HIP_ARCH=gfx942
export GLM53_BACKEND=gpu

The backend can also be selected automatically:

export GLM53_BACKEND=auto

Use the CPU backend explicitly with:

export GLM53_BACKEND=cpu

Documentation

Updated:

• CHANGELOG.md
• README.md
• GPU_BACKENDS.md
• docs/glm53-flash.md
• docs/cuda.md
• docs/benchmarks.md
• docs/ENVIRONMENT.md

Notes

Compiled test binaries and other generated ELF artifacts were intentionally excluded from this PR.