Add Lunar Lake Xe2 iGPU compatibility report and benchmarks#342
Open
MegaStood wants to merge 3 commits intointel:mainfrom
Open
Add Lunar Lake Xe2 iGPU compatibility report and benchmarks#342MegaStood wants to merge 3 commits intointel:mainfrom
MegaStood wants to merge 3 commits intointel:mainfrom
Conversation
Comprehensive testing of 7 LLM models on MSI Claw 8 AI+ handheld (Intel Core Ultra 7 258V, Arc 140V Xe2 iGPU, 32GB LPDDR5x shared memory). Models benchmarked (vLLM XPU backend): - openai/gpt-oss-20b MXFP4: 22.5 tok/s single-user (recommended default) - Qwen3.5-4B AutoRound INT4: 23.4 tok/s (best for multi-service) - Qwen3-8B AutoRound INT4: 18.6 tok/s - Qwen3.5-9B BF16/FP8/sym_int4: 5-14.7 tok/s - Qwen3.5-35B-A3B, GLM-4.7-flash, Qwen3-30B-A3B: all OOM Key findings: - ~13 GiB practical model ceiling on 32GB shared memory - Layer-by-layer weight processing (NOT 2x bulk) for AutoRound/sym_int4 - Prefix caching: 51x TTFT improvement at 16K context - 88,576 KV cache tokens at 0.7 util (~2.7x concurrent at 32K) - Cold prefill: ~807 tok/s (16K in 20.3s), decode: 15.6-22.5 tok/s Includes: - Complete running recipes (gpt-oss-20b, Qwen3.5-4B, ASR, TTS) - Tool calling + reasoning parser configuration - Memory budget tables for single/multi-service configurations - Standard benchmark script - Pre-built sym_int4 quantizer .so (12KB, from BigDL-core) - Draft comment for vLLM issue #30359
Replace old split "Environment Variables" / "vLLM Launch Flags" sections (which hardcoded --quantization int4 --max-model-len 8192) with a unified "Environment Variables and Launch Flags" section showing a general pattern derived from all tested models (4B, 8B, 9B dense, 20B MoE). Required vs optional flags documented in tables with "Tested With" column. https://claude.ai/code/session_01JyMJU94Dq32vYBGMoMJM34
Model weights now load successfully (16.16 GiB) with disk swap. Two issues remain: (1) IPEX marlin_shuffle_weight DEVICE_LOST during MoE warmup, (2) no MLA KV compression — vLLM stores full expanded KV at ~940 KB/token instead of compressed latent at ~53 KB/token (18x inflation). With proper MLA support, GLM-4.7-flash would fit 32K context in ~1.66 GiB KV cache. https://claude.ai/code/session_01JyMJU94Dq32vYBGMoMJM34
MegaStood
force-pushed
the
claude/upstream-lunar-lake-benchmarks-CB5w6
branch
from
634b09c to
8fa0c1c
Compare
6 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds comprehensive Lunar Lake (Arc 140V Xe2 iGPU) compatibility documentation
and benchmark results for running LLMs on 32GB shared-memory handhelds.
Tested On