Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
86 commits
Select commit Hold shift + click to select a range
1fcd662
Add Q4_K attention dispatch and direct GGUF requantization
GiorgioOppo Sep 9, 2026
c9474f6
Port automatic Q4 attention kernels across GPU backends
GiorgioOppo Sep 9, 2026
b8507c9
Document Q4 dispatch, SSD memory policy and validation
GiorgioOppo Sep 9, 2026
82fc226
Restore GPU decode defaults and fix ROCm KV ring writes
GiorgioOppo Sep 9, 2026
2c9d48f
Account for fixed weights in Metal manual SSD cache budgets
GiorgioOppo Sep 9, 2026
670e6b9
Add isolated CUDA Q4 grouped tok8 benchmark and validation
GiorgioOppo Sep 9, 2026
4a87d9c
Restore GB10 F16 compressor and aligned Q8 HC decode
GiorgioOppo Sep 9, 2026
080ac7d
Fuse ROCm F16 compressor stores and enable Q4 prefill at 8K
GiorgioOppo Sep 10, 2026
20e728e
Optimize Metal F16 compressor stores and enable Q4 prefill at 8K
GiorgioOppo Sep 10, 2026
2e61115
Optimize Metal M1 IQ2 sign decoding with popcount
GiorgioOppo Sep 10, 2026
2147a3e
Remove redundant IQ2 sign lookups from CUDA routed kernels
GiorgioOppo Sep 10, 2026
a081850
Remove redundant IQ2 sign lookups from ROCm routed kernels
GiorgioOppo Sep 10, 2026
0fe51ea
Optimize ROCm prefill routing and quantized kernel staging
GiorgioOppo Sep 10, 2026
b09f8e3
Reuse ROCm prefill activation preparation on gfx1151
GiorgioOppo Sep 10, 2026
d0bc6f0
Reuse Metal prefill activations for Q4 Q-B and IQ2 experts
GiorgioOppo Sep 10, 2026
0edc553
Separate GPU execution phases and reuse HC normalization operands
GiorgioOppo Sep 11, 2026
35da332
Optimize Metal indexer staging and exact top-k reduction
GiorgioOppo Sep 11, 2026
8a4ff9a
Add an isolated ROCm Q4 prefill lookahead candidate
GiorgioOppo Sep 11, 2026
809386d
Add prepared indexer operands and ROCm register-score candidates
GiorgioOppo Sep 11, 2026
7a5002d
Document GPU phase policies, indexer benchmarks and validation limits
GiorgioOppo Sep 11, 2026
a3043bb
Optimize V4.1 Metal attention and overlap SSD expert loading
GiorgioOppo Sep 12, 2026
2b9d1c2
Fuse V4.1 Metal BF16 boundaries in normalization and Q8 decode
GiorgioOppo Sep 12, 2026
8775288
Skip unused V4.1 index selection in full-KV batched prefill
GiorgioOppo Sep 12, 2026
a2e2ea5
Recover partial Metal SSD cache reservations on worker failure
GiorgioOppo Sep 12, 2026
bac91c2
Accelerate V4.1 Metal SSD prefill with explicit expert buffers
GiorgioOppo Sep 12, 2026
4792b5b
Add V4.1 Metal Q4 attention with FP16 prefill kernels
GiorgioOppo Sep 12, 2026
e4e74b4
Collect V4.1 attention imatrix inputs with release-safe validation
GiorgioOppo Sep 12, 2026
8cf2383
Add calibrated V4.1 Q4 attention conversion and APFS requantization
GiorgioOppo Sep 12, 2026
f6f036c
Fix CUDA long-context smoke test linking
Matthley Sep 12, 2026
a1c2574
Merge V4.1 Metal SSD and Q4 support into aprojq4-dense-attention
GiorgioOppo Sep 13, 2026
9f7e4eb
Fuse V4.1 shared Q8 projections and BF16 SwiGLU on Metal
GiorgioOppo Sep 13, 2026
73aae30
Reuse dead attention workspace for V4.1 Q4 Q-B FP16 matmul
GiorgioOppo Sep 13, 2026
4ec58cc
Cache V4.1 zero-prefix attention masks and block maps
GiorgioOppo Sep 13, 2026
7e593df
Dispatch V4.1 MoE prefill through N8 and N16 Metal tiles
GiorgioOppo Sep 13, 2026
2da41a4
Fuse V4.1 BF16 activation and hyper-connection epilogues
GiorgioOppo Sep 13, 2026
8cda676
Stage sixteen KV rows per V4.1 indexed prefill iteration
GiorgioOppo Sep 13, 2026
8ec5f79
Batch decoder KV publication and skip empty prefill submissions
GiorgioOppo Sep 13, 2026
48915fb
rocm: add DeepSeek V4.1 support for gfx1151
kyuz0 Sep 12, 2026
4f33140
docs: qualify 256K allocation with 64K ROCm context checks
kyuz0 Sep 13, 2026
249f627
Reuse packed Q2 weights and shared FP16 inputs in V4.1 Metal MoE
GiorgioOppo Sep 13, 2026
99de83c
Skip redundant V4.1 candidate ranking and retain growing top-k scratch
GiorgioOppo Sep 13, 2026
ecf0b13
Reuse selected V4.1 KV rows during sequential Metal decode
GiorgioOppo Sep 13, 2026
c12d639
Extend V4.1 SSD prefill sweeps and release temporary Metal storage
GiorgioOppo Sep 13, 2026
fb1516f
Isolate DeepSeek V4.1 ROCm allocation and graph paths
kyuz0 Sep 13, 2026
f120040
Align bulk attention references and document isolation regressions
kyuz0 Sep 13, 2026
c89c085
docs: describe final V4.1 ROCm support and validation
kyuz0 Sep 13, 2026
482898b
Reuse transient FP16 weights for V4.1 Metal Q4 output-B prefill
GiorgioOppo Sep 13, 2026
5d83b55
Fuse V4.1 scalar Metal decode HC and BF16 epilogues
GiorgioOppo Sep 13, 2026
a6c27dd
Extend V4.1 Metal MoE benchmarks across large prefill batches
GiorgioOppo Sep 13, 2026
3c346d0
docs: clarify Strix Halo GTT limits for SSD cache
kyuz0 Sep 13, 2026
8cc0c92
rocm: clarify V4.1 SSD cache fitting diagnostics
kyuz0 Sep 13, 2026
006b3c1
rocm: avoid extra GPU query in cache diagnostic
kyuz0 Sep 13, 2026
96b4659
Merge antirez/main V4.1 CUDA support into aprojq4-dense-attention
GiorgioOppo Sep 13, 2026
8d6f3cc
Batch V4.1 CUDA Q4 attention output with scalar reduction order
GiorgioOppo Sep 13, 2026
d0b8be6
feat(rocm): add two-rank V4.1 inference over TCP, USB4STREAM and RoCE
kyuz0 Sep 13, 2026
05d6991
docs: document V4.1 ROCm clustering setup and qualification
kyuz0 Sep 13, 2026
e35891f
docs(rocm): expose cluster options and measured transport performance
kyuz0 Sep 13, 2026
bcfa3ef
Merge upstream V4.1 CUDA support while preserving qualified ROCm tran…
kyuz0 Sep 14, 2026
2df985b
Merge antirez/main CUDA SSD prefill updates into aprojq4-dense-attention
GiorgioOppo Sep 14, 2026
2f45ce8
Fuse V4.1 CUDA Q4 output-A BF16 rounding into existing epilogues
GiorgioOppo Sep 14, 2026
7cb7944
Fuse V4.1 CUDA prefill HC expansion and BF16 boundaries
GiorgioOppo Sep 14, 2026
a3b530f
Fuse V4.1 Metal BF16 rounding and FP16 output-B RHS conversion
GiorgioOppo Sep 14, 2026
e651051
ROCm: use wave-oriented V4.1 scalar Q8 and indexer kernels
kyuz0 Sep 14, 2026
61eb12c
ROCm: accelerate V4.1 routed decode across resident, SSD and TP paths
kyuz0 Sep 14, 2026
9fdbb5a
ROCm: queue resident V4.1 layers and size TP reductions to payloads
kyuz0 Sep 14, 2026
eb45132
docs: record current V4.1 ROCm quality, SSD and cluster measurements
kyuz0 Sep 14, 2026
5cadfd5
Merge antirez/main Qwen3.8 support into aprojq4-dense-attention
GiorgioOppo Sep 14, 2026
27bc85b
Merge current main and qualify V4.1 ROCm performance integration
kyuz0 Sep 14, 2026
0fef357
docs: make the cluster worker visible to the benchmark workload watcher
kyuz0 Sep 14, 2026
a915971
docs: simplify ROCm setup and remove personal automation instructions
kyuz0 Sep 14, 2026
449a38d
rocm: accelerate V4.1 batched prefill on gfx1151
kyuz0 Sep 15, 2026
f24cd3a
rocm: overlap V4.1 shared experts with peer exchange
kyuz0 Sep 15, 2026
09f12d4
docs: record qualified ROCm prefill results and commands
kyuz0 Sep 15, 2026
cc27cae
rocm: partition V4.1 vocabulary rows between TP peers
kyuz0 Sep 15, 2026
85bd209
rocm: fuse V4.1 scalar TP wait and reduction
kyuz0 Sep 15, 2026
1e2e033
rocm: parallelize V4.1 decode attention across FP32 key tiles
kyuz0 Sep 15, 2026
c14f8cb
docs: record qualified ROCm decode results and reproduction
kyuz0 Sep 15, 2026
0ed0aa1
rocm: keep cluster transports to TCP and RoCE
kyuz0 Sep 16, 2026
40b1f63
Merge main and preserve ROCm model-cache call compatibility
kyuz0 Sep 16, 2026
8d5b2eb
Merge upstream main into DeepSeek V4.1 ROCm integration
kyuz0 Sep 17, 2026
7d0454b
QA: document V4.1 concurrency repeatability and retain full test evid…
kyuz0 Sep 17, 2026
8164e28
docs: remove concurrency issue report from server guide
kyuz0 Sep 17, 2026
9a73d66
Merge antirez/ds4#1036: integrate DeepSeek V4.1 ROCm support with Q4 …
GiorgioOppo Sep 19, 2026
5ef7727
feat(rocm): add typed Q4 attention output kernels for DeepSeek V4.1
GiorgioOppo Sep 19, 2026
a23e12b
feat(deepseek41): enable Q4 attention in ROCm graphs and TP2
GiorgioOppo Sep 19, 2026
eabe75a
test(rocm): cover V4.1 Q4 attention dispatch and numerical validation
GiorgioOppo Sep 19, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
65 changes: 65 additions & 0 deletions .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,8 @@
/tests/test_mxfp4_dot
/tests/test_mxfp4_metal
/tests/test_q4k_dot
/tests/test_cpu_q4_dense
/tests/test_quantizer_indexer_q4
/tests/test_sampling
/tests/test_web_recovery
/tests/test_quality_api
Expand All @@ -60,8 +62,16 @@
/tests/test_qwen4_vision
/tests/test_deepseek4_vision_image
/tests/test_deepseek41_metal
/tests/test_deepseek41_q8_bf16
/tests/test_deepseek41_shared
/tests/test_deepseek41_qb
/tests/test_deepseek41_outb
/tests/test_deepseek41_masks
/tests/test_deepseek41_cuda
/tests/test_cuda_q8_rows
/tests/test_cuda_v41_q4_output
/tests/test_cuda_v41_hc
/tests/test_deepseek41_bf16_rhs
/tests/test_cuda_reductions
/tests/test_cuda_shared
/tests/test_cuda_ssd_cache
Expand All @@ -71,6 +81,8 @@
/tests/test_qwen4_ngrams
/tests/test_qwen4_ngram_state
/tests/test_deepseek41_graph
/tests/test_deepseek41_q4_attention
/tests/test_deepseek41_imatrix_release
/tests/test_deepseek41_prefill
/tests/test_deepseek41_cli
/tests/test_engram
Expand All @@ -84,5 +96,58 @@ __pycache__/
.*.swp
.DS_Store

# Q4 test and benchmark build products
/tests/test_ssd_cache
/tests/test_metal_q4_hc
/tests/test_metal_decode_defaults
/tests/test_metal_f16_compressor
/tests/test_q4_epilogue_host
/tests/test_q4_prefill_dequant_host
/tests/test_cuda_q4_prefill_dequant
/tests/test_rocm_q4_prefill_dequant
/tests/test_q4_prefill_reduce_host
/tests/test_q4_prefill_reduce_host_fast
/tests/test_cuda_q4_prefill_reduce
/tests/test_rocm_q4_dot_host
/tests/test_rocm_q4_dot_host_fast
/tests/test_rocm_q4_lds_host
/tests/test_rocm_q4_lds_aligned_host
/tests/test_rocm_q4_lds_aligned_host_fast
/tests/test_rocm_q4_lds_aligned
/tests/test_rocm_q4_wmma_load_host
/tests/test_rocm_q4_qb_epilogue_host
/tests/test_rocm_q4_qb_epilogue_host_fast
/tests/test_rocm_q4_qb_epilogue
/tests/test_rocm_q4_dense_pair
/tests/test_metal_q4_prefill_pair
/tests/test_metal_indexer_q4
/tests/test_metal_q4_attn_out_a_direct
/tests/test_metal_q4_qb_f16_cache
/tests/test_metal_q4_qb_token_pair
/speed-bench/metal_q4_dense_pair_bench
/speed-bench/metal_q4_prefill_pair_bench
/speed-bench/metal_q4_mm_tail_cull_bench
/speed-bench/metal_q4_attn_out_a_direct_bench
/cuda/mmq/test/test_mmq_parity
/tests/test_cuda_q4_epilogue
/speed-bench/rocm_q4_prefill_bench
/speed-bench/cuda_q4_prefill_bench
tests/test_deepseek41_publication
tests/test_deepseek41_epilogues

/tests/test_deepseek41_candidates
/tests/test_deepseek41_gather
# Linux TP protocol and ROCm ownership harnesses
tests/test_tp_linux
ds4-kernel-v41-tp-attention
ds4-kernel-v41-tp-moe
ds4-kernel-v41-tp-gates

/ds4-kernel-v41-tp-mmq

/ds4-kernel-v41-tp-down

/ds4-kernel-v41-tp-bind-failure
/ds4-kernel-v41-q4-output
/tests/test_qwen4_conv_parallel
/tests/test_q8_prefill_variants
Loading