Skip to content

GPT-QModel v7.11.2

Choose a tag to compare

@Qubitium Qubitium released this 18 Nov 02:42
· 105 commits to main since this release
e75181d

GPT-QModel 7.11.2

  • Inference: optimize() now compiles the forward pass used by generation, and deleting compiled models releases VRAM. Supported CPU kernels compile with Inductor without graph breaks. EXL3 uses native kernels for BF16, and Paro selects Triton on ROCm.
  • Models: Added quantization for SenseNova-Loop 9B and Decision Transformer, which calibrates on trajectories without a tokenizer. Fixed checkpoint round trips for Qwen2.5-Omni, Qwen3-Omni, and InternVL.
  • Quantization: GPTQ and EXL3 use less GPU memory. NF12 and MACaQ can resume from checkpoints, and models stay runnable after failed or cancelled runs.
  • Calibration: Fixed token positions and masks in padded and multimodal batches. Chat datasets retain tool and document context, and AWQ still scales MoE experts that received tokens when others received none.
  • GGUF: Added ggml 1.0 support, including reading and dequantizing FP8 E4M3 tensors. The gguf extra requires llama-cpp-python>=0.3.37.
  • Compat: Fixed GPTQ and AWQ CPU offload in Transformers and added CPU readers for AWQ GEMV_FAST and LLM_AWQ. ParoQuant is now Paro; existing checkpoints still load.

API note: Use CalibrationConfig for calibration options. GPTQModel.export() and public MLX-LM conversion were removed.

Full list of changes

What's Changed

  • fix(awq): never fold a scale into an excluded projection (#4651) by @ZX-ModelCloud in #4659
  • Update version.py by @Qubitium in #4672
  • fix(calibration): use provided input_ids over a co-existing messages column (#4670) by @Qubitium in #4673
  • fix(granitemoehybrid): declare every checkpoint-owned decoder owner by @ZX-ModelCloud in #4674
  • fix(mlx): map per-layer-type RoPE onto mlx-lm's local/global pair (#4678) by @Qubitium in #4684
  • fix(config): reject bad bool values and unknown dynamic keys (#4668) by @Qubitium in #4693
  • fix(mlx): mirror the Torch affine grid fit and pack inference-mode-loaded weights by @Qubitium in #4687
  • fix(awq): reuse one clip-search scratch set across full tiles (#4682) by @Qubitium in #4696
  • fix(awq): never replay a recurrent mixer with the attention grid (#4671) by @Qubitium in #4694
  • fix(awq): load packed AWQ MoE experts through the native HF bridge (#4661) by @ZX-ModelCloud in #4692
  • fix(looper): stop claiming GPUs a quantization run excluded by @ZX-ModelCloud in #4688
  • fix(save): reject unknown save() keywords on the quantized path (#4686) by @Qubitium in #4698
  • fix(config): reject model-wide rotation on non-GPTQ formats (#4666) by @Qubitium in #4701
  • fix(hf): point refusals at a loader that can read the layout (#4676) by @Qubitium in #4700
  • fix(hf): forward quantization_config["backend"] through the load bridges (#4680) by @Qubitium in #4702
  • fix(looper): stop a subset only after every member captured (#4679) by @Qubitium in #4703
  • fix(hf): refuse GPT-QModel BitsAndBytes checkpoints in plain transformers (#4664) by @Qubitium in #4706
  • test(stage): give the stage doubles the in-flight layer hooks by @Qubitium in #4704
  • feat(calibration): exclude attention sinks in the shared calibration mask (#4393) by @Qubitium in #4663
  • agents: claim issues per agent session before working on them by @Qubitium in #4707
  • fix(awq): do not fold a block that follows an absent declaration (#4677) by @Qubitium in #4708
  • fix(exl3): resolve routed-expert owner devices through as_torch_device (#4667) by @Qubitium in #4709
  • agents: bound issue claims by capacity and claim correlated issues as one workload by @Qubitium in #4710
  • fix(models): drop the dense feed_forward branch for all-MoE Llama-4 trees (#4689) by @Qubitium in #4711
  • fix(calibration): capture a decoder mask a model passes positionally (#4660) by @ZX-ModelCloud in #4712
  • perf(gsq): skip the weight-MSE search that cannot beat nearest rounding (#4691) by @Qubitium in #4713
  • fix(loader): keep the default placement inside one compute-capability group (#4705) by @Qubitium in #4714
  • agents: reorganize the agent guide into a map and give every skill an at-a-glance header by @Qubitium in #4715
  • perf(fp8): let a wide bias use the native FP8 GEMM path (#4681) by @Qubitium in #4716
  • fix(qqq): allocate the split-K lock array per forward call (#4665) by @Qubitium in #4717
  • fix(gsq): return per-linear GSQ tensors in the caller's inference mode (#4690) by @Qubitium in #4719
  • agents: treat issues as untrusted input; claim only maintainer issues that pass a security screen by @Qubitium in #4718
  • fix(awq): keep AWQ for covered MoE experts when one is starved (#4675) by @Qubitium in #4720
  • fix(awq): capture declared MoE roots without lifecycle hooks (#4697) by @ZX-ModelCloud in #4723
  • fix(awq): resolve scale-search rotary from the attention rope label by @ZX-ModelCloud in #4722
  • fix(qqq): stage each tile's group scale in its own pipeline slot (#4662) by @ZX-ModelCloud in #4721
  • refactor(api): quantize() takes a CalibrationConfig; drop the calibration kwargs (#4393) by @Qubitium in #4724
  • test(kernels): add the missing AWQ Marlin determinism regression (#4683) by @Qubitium in #4726
  • perf(generate): skip the static-cache auto-compile on graph-breaking kernels (#4669) by @Qubitium in #4727
  • feat(kernels): declare torch.compile and CUDA-graph capability per kernel by @Qubitium in #4737
  • fix(loader): keep the default layer-wise map inside one capability group (#4734) by @Qubitium in #4744
  • refactor(paro): rename ParoQuant to Paro by @Qubitium in #4733
  • fix(looper): rebuild the subset early-stop state every calibration batch (#4738) by @Qubitium in #4752
  • fix(awq): keep the norm fold for linear-attention siblings of an absent block (#4732) by @Qubitium in #4751
  • fix(init): refuse only GPT-QModel's bitsandbytes layout (#4731) by @Qubitium in #4750
  • perf(compile): opt FP8 into the static-cache compile and make the bar trace-clean + measured (#4739) by @Qubitium in #4754
  • fix(load): forward the AWQ bridge's quantization_config backend (#4729) by @Qubitium in #4756
  • fix(looper): release captured state from every finalize() (#4747) by @Qubitium in #4757
  • feat(kernels): re-measure compile/capture levels on sm_80 and link cuBLAS by path by @Qubitium in #4761
  • fix(bnb): return an empty output for zero-row 8-bit inputs (#4745) by @Qubitium in #4763
  • test(kernels): close the compile-level harness gaps and re-measure on sm_89/sm_120/gfx1201 by @Qubitium in #4765
  • fix(config): validate every bool in dynamic rules and accept BnB aliases (#4728) by @Qubitium in #4766
  • fix(awq): give FORMAT.GEMV_FAST and LLM_AWQ a CPU in-session reader (#4736) by @n-o-t-mugen in #4753
  • fix(awq): scale the sibling branch of a shared norm instead of falling back (#4741) by @Qubitium in #4772
  • fix(load): install the Optimum GPTQ bridge regardless of import order (#4749) by @Qubitium in #4771
  • fix(kernels): measure BitsAndBytesLinear at CUDAGRAPH on sm_89/sm_120/gfx1201 by @Qubitium in #4773
  • test(kernels): classify the portable CPU readers in the fusion inventory (#4599, #4736) by @Qubitium in #4774
  • fix(bitsandbytes): fall back for misaligned INT8 cuBLASLt IMMA shapes (#4746) by @Qubitium in #4775
  • fix(qqq): pack the activation for column-strided inputs (#4740) by @Qubitium in #4776
  • fix(gguf): return an empty output for zero-row GGUF CUDA inputs (#4743) by @Qubitium in #4778
  • fix(gguf): return an empty output for M=0 on the CUDA kernel (#4743) by @Qubitium in #4780
  • test(hf): pin the import-order-independent Optimum GPTQ bridge (#4749) by @Qubitium in #4779
  • fix(looper): materialize non-persistent buffers before the resume hook (#4777) by @Qubitium in #4781
  • refactor(gguf): drop the duplicated M=0 guard in GGUFCudaKernel.forward by @Qubitium in #4782
  • perf(gguf): keep one resident copy of packed K-quant weights in the Triton cache (#4748) by @Qubitium in #4784
  • test: make the Hessian, free-threading and 2-GPU gates actually run (#4768) by @Qubitium in #4787
  • fix(awq): search the MoE root feature for shared-expert gate/up groups (#4760) by @Qubitium in #4790
  • docs: fix stale #4724 example, rotation kwarg, EoRA save path, and drifted contract refs (#4762) by @Qubitium in #4791
  • fix(qlinear): serialize state_dict with the CPU int4pack cold transform (#4788) by @Qubitium in #4798
  • fix(awq): keep the norm fold for an absent sibling branch (#4759) by @Qubitium in #4800
  • fix(exl3): publish the native handle under a per-module lock (#4793) by @Qubitium in #4803
  • fix(looper): make NF12 and MACaQ checkpoints resumable (#4792) by @Qubitium in #4806
  • fix(omni): save and load Qwen2.5-Omni checkpoints in the source HF key layout (#4758) by @Qubitium in #4810
  • perf(exl3): keep bf16 activations on the native fp16 kernels (#4767) by @Qubitium in #4817
  • fix(qqq): pack column-strided activations and guard the native A operand (#4740) by @Qubitium in #4804
  • fix(marlin): size the FP32 bias stage and keep it out of the reduction (#4735) by @Qubitium in #4819
  • fix(awq): keep the compiled decode dtype at the eager choice (#4794) by @Qubitium in #4822
  • fix(looper): close the Device-SMI handles on a failed or cancelled run (#4813) by @Qubitium in #4823
  • fix(export): refuse an hf export whose method has no dense decoder (#4786) by @Qubitium in #4824
  • fix(omni): resolve the Qwen3-Omni thinker embedding endpoints (#4816) by @Qubitium in #4825
  • fix(rocm): key the HIP JIT cache by the gfx target (#4815) by @Qubitium in #4826
  • fix(export): load MLX exports on CPU so GEMV checkpoints export on GPU hosts (#4809) by @Qubitium in #4828
  • fix(looper): keep the offload root for completed layers of a cancelled run (#4807) by @Qubitium in #4829
  • fix(mlx): run the target model's sanitize before loading packed weights (#4808) by @Qubitium in #4830
  • fix(paro): keep one packed CPU layout per layer (#4814) by @Qubitium in #4832
  • refactor(api): remove the deprecated GPTQModel.export() API by @Qubitium in #4831
  • fix(macaq): release the compiled flip sweep's frames per Linear (#4801) by @Qubitium in #4833
  • fix(exl3): reconstruct through the portable kernel when the trellis is off CUDA (#4789) by @Qubitium in #4835
  • test: realign stale FP8, Marlin, GGUF and decoder tests (#4764) by @Qubitium in #4836
  • perf(exl3): tile Hessian symmetrization and release the capture partial (#4818) by @Qubitium in #4837
  • test: stop the harness pinning PYTORCH_ALLOC_CONF over the backend default (#4821) by @Qubitium in #4838
  • fix(hf): install and repair the AWQ load bridge independent of import order (#4795) by @Qubitium in #4839
  • perf(macaq): keep the MXFP4 exponent screen off int64 planes (#4820) by @Qubitium in #4840
  • fix(gguf): order a cold Triton weight cache against the consuming stream (#4797) by @Qubitium in #4841
  • fix(cpu): detach the operands the int4 repack keeps from the checkpoint mmap (#4785) by @Qubitium in #4842
  • fix(cpu): rebuild the AVX512 checkpoint qweight instead of cloning it (#4783) by @Qubitium in #4845
  • perf(gsq): keep the ill-conditioned metric-factor fallback off fp64 GPU workspace (#4812) by @Qubitium in #4844
  • test: fix stale suite-triage doubles and oracles (#4769) by @Qubitium in #4846
  • fix(cpu): detach the AWQ int4 operands from the checkpoint mmap (#4785 follow-up) by @Qubitium in #4849
  • perf(gguf): read the packed Triton codes coalesced with one resident copy (#4811) by @Qubitium in #4851
  • fix(vl): map legacy checkpoint keys onto the export shell (#4802) by @Qubitium in #4853
  • fix(macaq): keep the compiled flip sweep graph-break free (#4801) by @Qubitium in #4854
  • fix(threadx): install the worker intra-op budget per worker thread (#4796) by @Qubitium in #4855
  • fix(qqq): replace the CPU checkpoint layout with the packed cache (#4805) by @Qubitium in #4856
  • fix(vl): keep an untied InternVL lm_head through lazy load and save (#4799) by @Qubitium in #4852
  • test: select CUDA devices relative to the host, not cuda:6/7/8 (#4768) by @Qubitium in #4857
  • fix(marlin): give each CUDA stream its own lock workspace (#4683) by @Qubitium in #4859
  • fix(mem): drop the reference cycles a single-device GPU load leaves behind (#4847) by @Qubitium in #4871
  • docs: re-anchor contract refs after writer.py growth (#4864) by @Qubitium in #4877
  • fix(cpu): copy a checkpoint operand only when it aliases the shard mapping (#4876) by @Qubitium in #4878
  • test(kernels): load through a fresh Lora so the dense lm_head is not targeted (#4764) by @Qubitium in #4886
  • fix(gguf): retire the Triton cache ready-event so CUDA graph capture works again (#4867) by @Qubitium in #4896
  • fix(load): make explicit CPU-offload device maps survive the first forward (#4866) by @Qubitium in #4890
  • fix(looper): materialize unreached layers after a cancelled run (#4860) by @Qubitium in #4893
  • feat(decision-transformer): tokenizer-optional load and tensor calibration inputs (#4755) by @Qubitium in #4894
  • fix(awq): replay MPT's MLP with the residual stream it requires (#4872) by @Qubitium in #4901
  • fix(calibration): accept an HF Dataset whatever display format it carries (#4882) by @Qubitium in #4902
  • fix(gptq): cap sink-row Hessian contributions when desc_act disables act_group_aware (#4742) by @Qubitium in #4909
  • test(interop): forward the round-trip fixture's inputs on the model device (#4908) by @Qubitium in #4911
  • fix(awq): resolve Laguna's rotary label from the model config (#4874) by @Qubitium in #4910
  • fix(load): read the checkpoint with the int device ordinal so cuda:0 stays untouched (#4863) by @Qubitium in #4915
  • fix(cpu): detach the Fused int4 operands from the checkpoint mapping (#4873) by @Qubitium in #4918
  • fix(calibration): pack before the per-row policies when concat_size is set (#4892) by @Qubitium in #4916
  • fix(looper): release pre-looper resources on a failed quantize (#4862) by @Qubitium in #4906
  • fix(generate): make optimize() compile the forward generate() runs (#4669) by @Qubitium in #4912
  • feat(sensenova): add SenseNova-Loop 9B (sensenova_loop) support (#4865) by @Qubitium in #4897
  • test(gpu): move the excluded-device gate onto CalibrationConfig (#4868) by @Qubitium in #4919
  • fix(load): let a quantize device sequence span GPU compute capabilities (#4875) by @Qubitium in #4920
  • fix(calibration): render a messages row with its tools and documents (#4858) by @Qubitium in #4927
  • test(load): declare require_load_tokenizer on the DummyQModel doubles (#4923) by @Qubitium in #4926
  • test(exl3): route the #4793 inner-handle burst back through the native stub (#4895) by @Qubitium in #4924
  • perf(paro): run the CPU int4 GEMM outside the process-wide transform lock (#4885) by @Qubitium in #4922
  • test(placement): make the faked GPU topology homogeneous on any host (#4905) by @Qubitium in #4929
  • fix(capture): derive position_ids from the mask for a padded calibration batch (#4861) by @Qubitium in #4930
  • perf(macaq): drop the per-Linear gc.collect() from the flip path (#4900) by @Qubitium in #4928
  • fix(gguf): publish the permuted packed weight only after its copy ran (#4879) by @Qubitium in #4921
  • fix(generate): treat the native EXL3 kernel as a graph-breaking kernel (#4917) by @Qubitium in #4932
  • fix(qwen3-omni): map runtime keys to the source HF checkpoint layout (#4881) by @Qubitium in #4931
  • fix(exl3): scale BF16 activations per row, not per tensor (#4870) by @Qubitium in #4938
  • test(calibration): cover real-model left-pad capture parity (#4861) by @Qubitium in #4944
  • fix(awq): pack Conv1D projections in the Linear layout (#4887) by @Qubitium in #4925
  • fix(load): keep a repaired AWQ module in the model's mode (#4907) by @Qubitium in #4949
  • fix(qqq): make the class-swapped CPU kernel independent of init state (#4940, #4942) by @Qubitium in #4951
  • fix(load): default an absent require_load_tokenizer to the text path (#4941) by @Qubitium in #4954
  • test(load): declare require_load_tokenizer on the last two loader doubles (#4941) by @Qubitium in #4957
  • perf(calibration): strip pre-padded rows before sorting and batching (#4889) by @Qubitium in #4946
  • fix(load): place a declared tied pair before the checkpoint tie (#4943) by @Qubitium in #4960
  • fix(capture): keep mRoPE position ids for multimodal calibration (#4933) by @Qubitium in #4961
  • fix(protocol): reject coerced booleans and unknown keys instead of dropping them (#4939) by @Qubitium in #4963
  • fix(exl3): pre-scale bf16 rows only near the FP16 limit (#4870) by @Qubitium in #4966
  • perf(cpu): verify the int4pack inverse once per layout instead of rebuilding every qweight (#4884) by @Qubitium in #4965
  • perf(gguf): retire the Triton cache handshake on every stream, not just consumers (#4956) by @Qubitium in #4968
  • test(docs): anchor the four drifting base.py citations by symbol (#4967) by @Qubitium in #4971
  • fix(config): validate smoother parameters at construction (#4953) by @Qubitium in #4970
  • fix(load): reject unknown kwargs and a dropped quantize_config on quantized loads (#4945) by @Qubitium in #4972
  • perf(gptq): resolve the desc_act sink-row cap without syncing per add_batch (#4950) by @Qubitium in #4976
  • test: realign seven stale doubles and oracles (#4969) by @Qubitium in #4978
  • fix(looper): rehome finished layers' kernels after a failed quantize (#4947) by @Qubitium in #4979
  • fix(hf): mark v2 qzeros runtime layout on bridge load (#4952) by @Qubitium in #4980
  • fix(load): never test an int-ordinal device map for truthiness (#4964) by @Qubitium in #4983
  • fix(looper): give every data-parallel replica its own subset early-stop state (#4904) by @Qubitium in #4982
  • fix(rotation): validate the calibration data before rotating the model (#4935) by @Qubitium in #4977
  • fix(models): compile the unbound forward so optimize() leaves no cycle (#4937) by @Qubitium in #4986
  • fix(fp8): expose the packed weight as qweight so PEFT LoRA can wrap it (#4948) by @Qubitium in #4987
  • perf(exl3): keep the bf16 pre-scale decision on the device, no host read per forward (#4955) by @Qubitium in #4989
  • fix(gguf): stop iterating the Triton cache dict on the forward path (#4962) by @Qubitium in #4988
  • fix(gptaq): give every calibrated processor the same shuffled batch order (#4869) by @Qubitium in #4985
  • test(exl3): run the EXL3 tests that were left permanently skipped (#4981) by @Qubitium in #4991
  • fix(paro): keep training gradients in the module forward (#4958) by @Qubitium in #4992
  • fix(vl): tie an InternVL head from the wrapper flag and checkpoint evidence (#4973) by @Qubitium in #4993
  • fix(calibration): keep messages conversations from numpy/pandas slices (#4936) by @Qubitium in #4994
  • fix(awq): pack token-aligned replay kwargs with the captured rows (#4880) by @Qubitium in #4996
  • test(coverage): fix the stale MoE sweep and the last nvcc gate (#4981) by @Qubitium in #4998
  • fix(awq): keep the dense-prefix norm group when the MoE root is empty (#4903) by @Qubitium in #5000
  • fix(awq): fold a norm's scale consumers outside the replay boundary (#4898) by @Qubitium in #5002
  • fix(qlinear): never return a reshape view from the trainable kernels (#4974) by @Qubitium in #5003
  • perf(load): remap legacy-layout safetensors keys without copying shards (#4959) by @Qubitium in #4999
  • fix(generate): stop forcing trace-clean kernels eager (#4984) by @Qubitium in #5006
  • fix(kernels): finish the view-free epilogues for wide bias and AWQ GEMM (#4974) by @Qubitium in #5001
  • fix(load): accept the Hub staging options a quantized load forwards (#5007) by @Qubitium in #5016
  • fix(calibration): render Hub tools columns the way inference does (#4934) by @Qubitium in #4997
  • fix(hymba): materialize root-level bare parameters and honour construction dtypes (#4888) by @Qubitium in #5021
  • fix(config): reject fraction asymmetric percentiles and validate dynamic smoothers (#5005) by @Qubitium in #5017
  • fix(gguf): build the Triton cache once per module so a cold burst cannot drop it (#5008) by @Qubitium in #5028
  • fix(awq): map kept-row indices back onto the padded grid (#5015) by @Qubitium in #5030
  • fix(awq): replay a **kwargs block and fold the norm the layer owns (#4891) by @Qubitium in #5019
  • fix(qqq): publish dropped checkpoint buffers before an export and fail closed on load (#5010) by @Qubitium in #5032
  • fix(exl3): quantize a zero-token routed expert with a data-free Hessian (#4995) by @Qubitium in #5025
  • fix(looper): leave a cancelled quantize() with a runnable model (#5009, #5012) by @Qubitium in #5031
  • perf(cpu): drop the int4 qweight snapshot where the inverse is exact (#5013) by @Qubitium in #5033
  • fix(gguf): keep the K-quant shift tables on the device so CUDA-graph capture works (#5014) by @Qubitium in #5035
  • fix(awq): clear the int4pack training latch in TorchFusedAwqLinear (#5018) by @Qubitium in #5036
  • fix(looper): return freed AWQ host pages at each layer boundary (#4990) by @Qubitium in #5037
  • fix(decision-transformer): cast tensor calibration rows to the model dtype by @Lumea-art in #5038
  • fix(looper): keep a post-loop tail failure inside the failed-run cleanup (#5023) by @Qubitium in #5039
  • fix(gptq): stop pinning the task Hessian across quantize() (#5026) by @Qubitium in #5040
  • fix(weight-only): share the tied-head check with the module looper (#5027) by @Qubitium in #5041
  • fix(exl3): republish only the native tensors the loaded model owns (#5029) by @Qubitium in #5042
  • perf(vptq): pack the index bitstream with word arithmetic instead of one int64 per bit (#5034) by @Qubitium in #5044
  • fix(gptq): stop holding the task Hessian across g.quantize() (#5026) by @Qubitium in #5043
  • fix(kernels): a per-format capture refinement may only narrow SUPPORTS_GRAPH_CAPTURE (#5014) by @Qubitium in #5045
  • fix(load): honour a CPU-offload device map in the HF bridge (#5022) by @Qubitium in #5046
  • fix(looper): drop a leftover del g_h that fails the ruff gate (#5026) by @Qubitium in #5047
  • fix(weight-only): return the in-flight layer to its home device on failure (#5011) by @Qubitium in #5049
  • test(load): keep the #5022 bridge nodes runnable without Optimum or CUDA (#5022) by @Qubitium in #5051
  • ci(awq-cpu): run the lane when the HF bridge regression file changes (#5022) by @Qubitium in #5052
  • fix(paro): pick the native Triton kernel on ROCm AUTO and make the dense fallback graph-capturable (#4914) by @Qubitium in #5053
  • fix(gptq): drop the stale del g_h that broke main's lint gate by @Qubitium in #5048
  • fix(gptq): harden the failed-quantize Hessian release and run the #5026 guards by @Qubitium in #5054
  • test: skip CUDA-only kernels on ROCm with a shared requires_nvidia_cuda gate (#4770) by @Qubitium in #5056
  • Require isolated agent workspaces and draft-first PRs by @Qubitium in #5057
  • test(marlin): derive the shared-memory budget invariant from the kernel sources (#4735) by @Qubitium in #5058
  • test: fix two stale #4769 section A items (bits_new NATIVE_MODEL_ID, jamba mamba_ssm gate) by @Qubitium in #5059
  • feat(gguf): support ggml 1.0 by @Qubitium in #5055
  • Teach the regression skill to check row identity, placement, run exits, and CI reachability by @Qubitium in #5061
  • Streamline the agent guide and all 13 repository skills by @Qubitium in #5062
  • fix(optimize): release dynamo's trace cycles so a compiled model frees its VRAM (#5020) by @Qubitium in #5064
  • test: realign residual suite triage with current kernels and layouts (#4769) by @Qubitium in #5063
  • fix(qqq): chunk CPU INT8 GEMM contractions past the INT32 limit by @ZX-ModelCloud in #5065
  • refactor(looper): give both quantization loops one run-exit recovery owner (#5060) by @Qubitium in #5066
  • Summarize GPT-QModel 7.11.1 in the README by @Qubitium in #5069
  • Teach four regression checks from the weekly fix audit by @Qubitium in #5071
  • perf(cpu): reach torch.compile/inductor on CPU quantized kernels by @Qubitium in #5070
  • fix(gguf): keep representable Q8_K subnormal scales by @ZX-ModelCloud in #5068
  • Mention CPU Inductor support in the 7.11.1 highlights by @Qubitium in #5074
  • perf(cpu): make NF12, Paro and QQQ inductor-ready on CPU by @Qubitium in #5075
  • refactor(mlx)!: remove the public convert_gptq_to_mlx_weights() API by @Qubitium in #5076
  • fix(vptq): keep the weighted mean of buckets with tiny positive weights by @ZX-ModelCloud in #5073
  • perf(cpu): make TorchQuantEmbeddings inductor-ready on CPU by @Qubitium in #5077
  • Streamline the 7.11.1 README highlights by @Qubitium in #5078
  • Update version.py by @Qubitium in #5079
  • Update README.md by @Qubitium in #5080

New Contributors

Full Changelog: v7.11.0...v7.11.2