Upgrade llama.cpp from b9094 to b9102

[bernardladenthin]

Summary

This PR upgrades the pinned llama.cpp version from b9094 to b9102, incorporating upstream improvements to CUDA AllReduce pipelines, SYCL im2col support, sampling fixes, and security updates.

Key Changes

• CUDA AllReduce Pipeline Refactoring: New ggml_cuda_ar_pipeline struct and APIs supporting 2-GPU PCIe AllReduce without NCCL (Volta+ / sm70+), with configurable chunked kernel vs copy-engine paths and environment variable tuning
• SYCL im2col Support: Added ggml_sycl_im2col_3d function enabling GGML_OP_IM2COL_3D on Intel GPUs with tile-based thread decomposition
• Sampling Bug Fix: Fixed common_sampler_sample to call set_logits before backend-sampling check and properly scan all tokens in cur_p.data, enabling correct post-sampling probabilities with backend sampling
• Model Loading Refactor: Moved n_vocab loading from llama_model_base::load_hparams() to per-model load_arch_hparams() implementations
• Security Fix: Updated cpp-httplib (0.43.4) with improved chunk-size parsing to prevent overflow and reject invalid chunk extensions
• GGML Version Bump: Patch version incremented from 0.11.0 to 0.11.1

Notable Details

• CUDA AllReduce supports platform-specific defaults (Linux→NCCL, Windows→internal) with GGML_CUDA_ALLREDUCE environment variable override
• Backend sampling now works correctly with post_sampling_probs and filters 0.0-probability tokens from results
• HIP/MUSA builds return nullptr stubs for AllReduce pipeline
• All changes are internal to llama.cpp; no JNI layer modifications required

https://claude.ai/code/session_01QopdxqEvbkhiaaBRqBzgzc

[claude]

🔍 PR Review: llama.cpp Upgrade b9094 → b9102

❌ CRITICAL ISSUE: CLAUDE.md Version Not Updated

The PR shows a diff updating CLAUDE.md, but the file was not actually modified:

Current state:

• Line 9 in CLAUDE.md still shows: Current llama.cpp pinned version: **b9094**
• Changelog entries for the b9094–b9102 range are missing

Expected:

• Line 9 should be: Current llama.cpp pinned version: **b9102**
• 7 new changelog entries should be added (CUDA AllReduce pipeline, SYCL im2col, sampling bug fix, server-context changes, model loading refactor, cpp-httplib security fix, ggml version bump)

✅ CORRECT: Version Updates in Other Files

• CMakeLists.txt (line 100): Correctly pinned to b9102 ✓
• README.md (line 2): Correctly updated badge to b9102 ✓

📋 Required Changes

Please update CLAUDE.md:

1. Change line 9 version from b9094 to b9102
2. Add the 7 new changelog table entries documented in the PR description

This is straightforward to fix — the detailed changelog text already appears in the PR description.

[claude]

⚠️ This line still shows b9094 but should be updated to b9102 to match the upgrade in CMakeLists.txt and README.md.

Suggested change

	Current llama.cpp pinned version: **b9102**
	Current llama.cpp pinned version: **b9102**

[claude]

📌 Missing changelog entries for b9094–b9102 range. According to the PR description, 7 new changelog rows should be added here:

1. CUDA AllReduce pipeline refactoring (ggml/src/ggml-cuda/allreduce.cu + allreduce.cuh)
2. CUDA multi-GPU changes (ggml/src/ggml-cuda/ggml-cuda.cu)
3. SYCL im2col support (ggml/src/ggml-sycl/ggml-sycl.cpp + im2col.cpp + im2col.hpp)
4. GGML version patch bump (0.11.0 → 0.11.1)
5. Sampling bug fix (common/sampling.cpp)
6. Server context changes (tools/server/server-context.cpp)
7. Model loading refactor (src/llama-model.cpp)
8. cpp-httplib security fix (vendor/cpp-httplib/)

These details are in the PR description and should be preserved in CLAUDE.md for future reference.

[claude]

✅ Correct: Version pinned to b9102