integration: bring heterogeneous GGUF cache into Qwen3.8 branch (conflict probe, closed) - #1
Closed
tmielsch wants to merge 4 commits into
Closed
Conversation
Adds the laguna GGUF architecture: hybrid full/SWA attention with per-layer query-head counts, QK RMSNorm, per-layer-type rope (YaRN partial-dim on full layers, plain rope on SWA), a per-head softplus attention output gate, and sigmoid + score-correction-bias MoE routing with one always-on shared expert. Semantics follow llama.cpp src/models/laguna.cpp. Poolside/Unsloth laguna checkpoints quantize per tensor, so this also generalizes the GGUF plumbing: - six more ggml types (Q3_K, Q4_K, Q5_K, IQ1_S, IQ2_S, IQ2_XXS, IQ3_XXS, IQ4_XS) wired into the dequant tables and mmvq/mmq/dequant dispatch - a "gguf" expert-bank format whose per-layer quant types vary: flat padded [E, stride] host banks plus a new expert_stride_bytes argument threaded through the vendored moe_vec launchers (0 = previous dense behaviour) - moe_vec calls chunked to min(65535, 16384) rows: the kernel indexes experts via blockIdx.z (CUDA grid-z cap 65535) and a 16k-token x top-8 prefill chunk also overflowed transient VRAM - q/k/v kept as separate projections (a layer may quantize attn_v differently from attn_q/k -- observed on the XS Q4_K_M build) - deferred GGUF linears materialized from the file's tensor table at conversion time, before the engine collects the state dict Tested on: RTX 5080 (16 GB), 23 GB RAM, NVIDIA 610.62, CUDA 13, wsl2. Checkpoints: unsloth/Laguna-S-2.1-GGUF (Laguna-S-2.1-UD-IQ1_S.gguf, S, metadata/tensor coverage only) and poolside/Laguna-XS-2.1-GGUF (Laguna-XS-2.1-Q4_K_M.gguf) plus a third-party APEX-I-Mini XS build (Q3_K/Q4_K/Q5_K/Q6_K/IQ2_S) run end to end. Validation on Laguna-XS-2.1-APEX-I-Mini.gguf: ft serve --model <model> --kv-cache-dtype fp8_e4m3 --num-tokens 262144 --kv-reserve-tokens 262144 - NIAH 3/3 at 250,054 tokens (needle at 10%/50%/90% depth, exact passcode), ~433 tok/s prefill - decode 157-162 tok/s at 64k ctx, 21-23 tok/s at ~250k ctx (PCIe-bound) Not done: S-model e2e (host lacked RAM for its ~37 GiB expert banks), hybrid/cpu MoE backend for the gguf format, FTW conversion, TP>1. See tasks/laguna-handover.md. Co-Authored-By: Claude <noreply@anthropic.com>
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
tmielsch
changed the title
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.
Temporary conflict-probe PR only. GitHub reported the expected conflicts between upstream FlashML-org#232 and the FlashML-org#102/FlashML-org#199 stack, so it was closed without merge. The compatible generic GGUF/geometry-cache blobs are being ported selectively into qwen38-flash-next-gguf instead.