Repository navigation

DeepSeek-V4.1-1017: experimental TCP MoE partitions and parallel attention - #1061

Closed

cwschroeder wants to merge 184 commits into

ashhart:mainfrom

cwschroeder:feat/ds1017-tp2-integration

Closed

DeepSeek-V4.1-1017: experimental TCP MoE partitions and parallel attention#1061
cwschroeder wants to merge 184 commits into
ashhart:mainfrom
cwschroeder:feat/ds1017-tp2-integration

Commits on Oct 18, 2026
- Restore shared Living Weights and reviewed lesson safety
  
  committed
  9e7f502
- Restore CUDA lane rounds with the reviewed decode fixes
  
  ashhart
  and
  BobClawblaw
  committed
  1892084
- Restore shared Flash Next vision with reviewed grammar and cache fixes
  
  committed
  bd5dbfb
- Restore per-engine Flash Next windows and reviewed rotary contracts
  
  ashhart
  and
  BobClawblaw
  committed
  209ee73
- Place live replay beside its lesson owner
  ashhart
  committed
  79af9d6
- Match conversation teaching and checks to reader capabilities
  ashhart
  committed
  dbd5b2f
- Keep lesson progress alive through long units
  ashhart
  committed
  b2338a8
- Keep dense learner reads separate from served graph variants
  
  ashhart
  and
  BobClawblaw
  committed
  88c3443
- Batch independent cold teacher prompts without changing lesson order
  ashhart
  committed
  b7e1b27
- Publish reading progress before the first long learner unit
  ashhart
  committed
  d3509f6
- Finish shard saving before closing the lesson stream
  ashhart
  committed
  c316b50
- Size captured Qwen launches for live rows and windows
  ashhart
  committed
  3238979
- Share startup warmups while measuring each decode width
  ashhart
  committed
  9db03c0
- Integrate runtime sessions and common EXL3 ownership
  
  committed
  cc85750
- Admit DeepSeek shared configuration and checkpoint layouts
  
  committed
  8fcad83
- Add DeepSeek Engram ownership and token maps
  
  committed
  3d75343
- Wire common EXL3 CUDA forms and launch contracts
  
  committed
  f755f4b
- Emit the DeepSeek block specialization inventory
  
  committed
  127b3ee
- Load DeepSeek runtime forms with checked staging
  
  committed
  335aca3
- Add paged split KV and shared session transactions
  
  committed
  8eccc8c
- Run DeepSeek forward and prefill through common state
  
  committed
  6d33a53
- Draft DeepSeek replies with keyed shared lane sampling
  
  committed
  1c80298
- Serve DeepSeek with owned device drafts and sessions
  
  committed
  b0f4bd8
- cuda: retain shared EXL3 resolvers and training image exports
  
  committed
  1964746
- cuda: retain lane forwarding and unsupported policy refusal
  
  committed
  eabe96c
- Install selected contract images through their current common owners
  
  committed
  0fe59c4
- Require explicit x3ld specialization on every selected architecture
  
  committed
  2c2d058
- Stabilize Engram and MiMo allocation-failure controls
  
  committed
  d2d63ca
- Register tree-round launches with repaired pending fixtures
  
  ashhart
  and
  jschmied
  committed
  283d2e5
- Extend row contracts with the reviewed rounding and pending declarations
  
  ashhart
  and
  jschmied
  committed
  b44ee9f
- Serve Kolibri through the current native and shared operator owners
  
  ashhart
  and
  jschmied
  committed
  7162134
- Batch fresh GLM prompts on the maintained rank owners
  
  ashhart
  and
  BobClawblaw
  committed
  76765b5
- Keep fixed-K GLM prompts explicit and disabled by default
  
  ashhart
  and
  BobClawblaw
  committed
  09d39bd
- Score GLM prompt rows with startup-owned output mappings
  
  ashhart
  and
  Bizuayeu
  committed
  9443d56
- Support GLM grammars with sequence-owned redraws and pending sampling contracts
  
  ashhart
  and
  Bizuayeu
  committed
  2b6c7a6
- Parse the GLM tool markup through maintained protocol owners
  
  ashhart
  and
  Bizuayeu
  committed
  d267438
- Version benchmark natural stops with guarded publication commands
  
  committed
  d2b6470
- Route chat-only benchmark cells from explicit capabilities
  
  committed
  5841619
- Declare the DeepSeek native family chat-only
  
  ashhart
  and
  BobClawblaw
  committed
  99b6572
- Split current Flash Next replay along its ownership seams
  ashhart
  committed
  a4bd65b
- Split current Flash Next loading and served generation
  ashhart
  committed
  ebb6bdc
- Split Metal build construction with the current candidate test wiring
  ashhart
  committed
  fc8502e
- Pin serving guards to the reviewed lane and decode baseline
  ashhart
  committed
  105398c
- Carry the runtime-one AOT fallback after specialized candidates
  
  committed
  3af275a
- Preserve nospec selection and capture checks before runtime-one fallback
  
  committed
  fc6d416
- Keep restored catalog and native owners within source limits
  ashhart
  committed
  f69c7fa
- Keep the vision refusal fixture aligned with the shared parser
  ashhart
  committed
  bb25309
- Avoid gathering already contiguous lone drafter mask rows
  ashhart
  committed
  45bb19a
- Build straight CUDA window schedules without the branch planner
  ashhart
  committed
  1eb81e5
- Propagate rotary probe errors through FP32 phases and BF16 endpoints
  
  committed
  2090adc
- Synchronize the halt fixture with the observed halt request
  ashhart
  committed
  c59e828
- Allocate DeltaNet rollback banks for the slots that need them
  ashhart
  committed
  7520d4c
- Restore EXL3 unpack helpers and check CUDA source includes
  ashhart
  committed
  1fe556f
- Price lone CUDA windows from their own complete rounds
  ashhart
  committed
  8aa29ba
- Restrict DeepSeek images to their supported SM targets
  ashhart
  committed
  3ad0497
- Report request times on completion lines
  ashhart
  committed
  86567eb
- Draw fresh sampling seeds for native requests
  ashhart
  committed
  8529473
- Add the host block-CSF scale codec (#1021)
  
  Bizuayeu
  authored and
  ashhart
  committed
  443cf3d
- Count indexed checkpoint shards and refuse incomplete estimates (#1022)
  
  Bizuayeu
  authored and
  ashhart
  committed
  6630a1c
- Describe per-module mixed-precision checkpoint recipes (#1023)
  
  Bizuayeu
  authored and
  ashhart
  committed
  836e952
- Pack FP8 scales from global rank-share origins (#1024)
  
  authored and
  ashhart
  committed
  9e423aa
- Read zstd-compressed CUDA resource images
  ashhart
  committed
  9b78f7b
- Fix DeepSeek launch bounds and bind every contract image
  ashhart
  committed
  caac51d
- Share completed CUDA shape warmup passes
  ashhart
  committed
  fb962e2
- Complete CUDA window calibration before starting the lane worker
  ashhart
  committed
  34ad2ed
- Size drafter scratch for served streams and report startup phases
  ashhart
  committed
  7e652ea
- Cache Qwen CUDA window costs before serving
  ashhart
  committed
  9c11d9d
- fix(cuda): preserve first-token handoff through native host
  
  jrb-ai
  authored and
  ashhart
  committed
  73d0881
- docs: update contributor links to jrb-ai
  
  jrb-ai
  authored and
  ashhart
  committed
  3a9b4d8
- Preserve the first-token policy through recorded backends
  ashhart
  committed
  36188d9
- Reuse affine-4 weights across 32 GB10 verify rows
  ashhart
  committed
  4d1a3f1
- Describe the composed 1.1.1 release changes
  ashhart
  committed
  6b5549a
- Report the source release version and guard changelog drift
  ashhart
  committed
  eb25143
- Require API keys on every private HTTP route
  ashhart
  committed
  788869d
- Reject HTTP requests from untrusted hosts and origins
  ashhart
  committed
  9fe5ecf
- Document efficiency-core pinning for long CUDA runs
  ashhart
  committed
  9f61b36
- Name optional runtime libraries and report their loader availability
  ashhart
  committed
  f70c4b9
- Select precise rotary math and qualify it by complete Metal sweeps
  
  committed
  9c8ce4b
- Make the optional Neural Engine prompt share opt-in
  ashhart
  committed
  4163a8b
- Document precise Metal rotary math and the GLM NVFP4 regression
  ashhart
  committed
  ad87a2d
- Gate experimental DeepSeek CUDA serving behind an explicit opt-in
  ashhart
  committed
  5b2e274
- Explain the GLM short-prompt grouping change in the release notes
  ashhart
  committed
  c9ff37b
- Render the HIP follower dependency refusal once
  ashhart
  committed
  72baf2e
- Refresh explicit serving approvals after verification and runtime fixes
  ashhart
  committed
  2cfa68c
- Keep Hub fake-IP checks and downloads on one HTTPS endpoint
  ashhart
  committed
  9fb530e
- Require a complete drive path shape in release payload text
  ashhart
  committed
  e9ee1aa
- Update contributor credit for renamed GitHub account
  ashhart
  committed
  574a7fe
- flashnext: restore the CUDA column copy and check literal kernel lookups
  
  ashhart
  and
  ZackSample
  committed
  77152dc
- qwen CUDA: keep prompt expert arithmetic independent of verify capacity
  ashhart
  committed
  3e3f895
- Refuse unavailable DeepSeek release backends consistently
  ashhart
  committed
  d2f079b
Commits on Oct 19, 2026
- DeepSeek V4.1: the Engram token map built in Zig, byte-identical to the release's
  
  akol1
  authored and
  cwschroeder
  committed
  8d4faf5
- DeepSeek V4.1 Flash: configuration, Engram tables and resident weight staging
  
  akol1
  authored and
  cwschroeder
  committed
  e16ec0f
- DeepSeek V4.1 Flash: Metal operators with fp64-dot references and the qualification runner
  
  akol1
  authored and
  cwschroeder
  committed
  31da5b2
- THIRD_PARTY_NOTICES: the native DS4.1 reference follows DeepSeek-V4.1-Flash's MIT inference
  
  akol1
  authored and
  cwschroeder
  committed
  60a16a3
- DeepSeek V4.1 load: a shard contributes only the tensors the index maps to it
  
  akol1
  authored and
  cwschroeder
  committed
  0dcf49d
- DeepSeek V4.1 load: the router gate and bias ride the pack's dtypes
  
  akol1
  authored and
  cwschroeder
  committed
  466a4c4
- DeepSeek V4.1 load: the Engram gate weights ride f32 or bf16
  
  akol1
  authored and
  cwschroeder
  committed
  3a7d6bb
- DeepSeek V4.1: the kept-dense EXL3 form, for rank-split repacks
  
  akol1
  authored and
  cwschroeder
  committed
  c69533f
- DeepSeek V4.1 reference: the kept-dense apply
  
  akol1
  authored and
  cwschroeder
  committed
  a942700
- DeepSeek V4.1 load: the shard-plan tests pass the index-ownership argument, and the duplicate-tensor case stops logging at error level
  
  akol1
  authored and
  cwschroeder
  committed
  197e25b
- DeepSeek V4.1: the rebase residue — one counters binding, the state-window kernel's brace
  
  akol1
  authored and
  cwschroeder
  committed
  5423bf0
- mHC fp64 ideal reference: the mix at rounding level, 1-ULP noise attenuated
  
  akol1
  authored and
  cwschroeder
  committed
  5805f65
- CSA2 reference: FP64-ideal parameterization of the pool, RMS and sparse-head boundaries
  
  akol1
  authored and
  cwschroeder
  committed
  d0af92e
- csa2: FP64 stage harness, compressor pool + RMS, six adversarial input families
  
  akol1
  authored and
  cwschroeder
  committed
  178dbec
- csa2: FP64 stage harness over the sparse attention kernel, with kernel-faithful key addressing
  
  akol1
  authored and
  cwschroeder
  committed
  f106e35
- csa2: indexer stage harness — FP32/FP64 score comparison and top-k selection agreement
  
  akol1
  authored and
  cwschroeder
  committed
  bac57d2
- csa2: indexer stage receipts — carry the BF16 einsum/score boundaries in the FP64 ideal, read positions through the view offset
  
  akol1
  authored and
  cwschroeder
  committed
  d5cebed
- csa2: sparse stage reads its own FP32 rounding level before the BF16 output quantum
  
  akol1
  authored and
  cwschroeder
  committed
  53e49d1
- csa2: per-stage harnesses for sinks, tile denominators, output mix and RoPE
  
  akol1
  authored and
  cwschroeder
  committed
  e4347ee
- csa2: measure how much score noise the top-k selection absorbs
  
  akol1
  authored and
  cwschroeder
  committed
  34d6746
- The fp64 harness ports: dedupe the shared math helpers, fix the shard-plan test calls
  
  akol1
  authored and
  cwschroeder
  committed
  d7aeb04
- The oracle records the indexer's selection margins while it teaches
  
  akol1
  authored and
  cwschroeder
  committed
  054a4ee
- The indexer-margin capture at depth: one diagnostic token's selection scratch, verbatim
  
  akol1
  authored and
  cwschroeder
  committed
  a3b31df
- DeepSeek V4.1: lane slots, each with its own caches and Engram cursor
  
  akol1
  authored and
  cwschroeder
  committed
  361548e
- tf-dsv41-run: the counters read the full rusage_info_v4
  
  akol1
  authored and
  cwschroeder
  committed
  d5347cb
- DeepSeek V4.1: the lane backend, a stream a slot
  
  akol1
  authored and
  cwschroeder
  committed
  7f3d4f2
- DeepSeek V4.1: the lane host behind the native server
  
  akol1
  authored and
  cwschroeder
  committed
  033bcaa
- serve: --engram-table-dir reaches the native engine
  
  akol1
  authored and
  cwschroeder
  committed
  2c839f8
- DeepSeek V4.1: the lane host builds the token map when it is missing
  
  akol1
  authored and
  cwschroeder
  committed
  51c76e6
- DeepSeek V4.1: the family guide
  
  akol1
  authored and
  cwschroeder
  committed
  3fd5b68
- the generated table through zig fmt
  
  akol1
  authored and
  cwschroeder
  committed
  3b8fbb6
- the operator harnesses through zig fmt
  
  akol1
  authored and
  cwschroeder
  committed
  1cbe4b8
- deepseek-v41: engram_rows, uncached row reads into compact staging
  cwschroeder
  committed
  ed35475
- deepseek-v41: --engram-stream, tables read per chunk instead of wired
  cwschroeder
  committed
  2f8947c
- deepseek-v41: engram_rows misses go through a Fetcher interface
  cwschroeder
  committed
  e57eb9d
- server: DeepSeek V4.1's DSML tool calls (a space after the bar)
  cwschroeder
  committed
  185caa5
- deepseek-v41: tf-dsv41-template-check, a template against a release's goldens
  cwschroeder
  committed
  1b28f74
- docs: DeepSeek V4.1 on 256 GB with --engram-stream
  cwschroeder
  committed
  9645c69
- tf-dsv41-run: lanes accepts its CONTEXT argument
  cwschroeder
  committed
  900cedd
- deepseek-v41: dense dots one column per SIMD group, loads issued ahead
  cwschroeder
  committed
  72f8875
- exl3: a two-column matmul for products too small to fill the GPU
  cwschroeder
  committed
  f346d02
- deepseek-v41: the eight wo_a slices as one indexed product
  cwschroeder
  committed
  c67e37a
- deepseek-v41: the shared expert as each token's seventh routed row
  cwschroeder
  committed
  c5b41cd
- deepseek-v41: batch wo_a and merge the shared expert only for few rows
  cwschroeder
  committed
  550b8fc
- exl3: encodeJobs, independent products as one concurrent group
  cwschroeder
  committed
  7ca779a
- deepseek-v41: a concurrent forward encoder with independent product groups
  cwschroeder
  committed
  3dbe151
- exl3: borrow output rotation blocks for Metal tensor parallelism
  cwschroeder
  committed
  ad6ab78
- exl3: partition resident expert descriptors by output blocks
  cwschroeder
  committed
  cc387e3
- deepseek-v41: add experimental two-rank MoE output partitions
  cwschroeder
  committed
  968f35f
- exl3: compare partitioned expert descriptors with full Metal outputs
  cwschroeder
  committed
  cb66809
- deepseek-v41: test repeated TP exchanges and poisoned peer channels
  cwschroeder
  committed
  b7ea8d2
- deepseek-v41: adapt streamed loading and tokenizer wiring to current main
  cwschroeder
  committed
  426efd1
- deepseek-v41: mirror native lane commands across the TP pair
  cwschroeder
  committed
  c3985f1
- exl3: stage the GPU qualification binary for remote hardware checks
  cwschroeder
  committed
  c9b2f9a
- deepseek-v41: add bounded TCP window transport for the TP pair
  cwschroeder
  committed
  9aecfae
- deepseek-v41: bound startup pairing separately from TCP message waits
  cwschroeder
  committed
  0f44e3e
- deepseek-v41: stage qualification binaries for protected GPU windows
  cwschroeder
  committed
  e18c7cb
- deepseek-v41: place transport comments beside their declarations
  cwschroeder
  committed
  25926d3
- Split EXL3 and DeepSeek numerical qualification by operator boundary
  cwschroeder
  committed
  1f1e4e2
- Record teacher wall time including tensor-parallel communication
  cwschroeder
  committed
  a86286f
- Make Engram Unicode tables reproducible and apply the generated-data length rule
  cwschroeder
  committed
  40d9972
- docs: preserve token map provenance and normalization caveat
  cwschroeder
  committed
  2d0079b
- docs: condense DeepSeek operator and loader comments
  cwschroeder
  committed
  53071ae
- docs: locate input projection contracts beside their declarations
  cwschroeder
  committed
  c66964b
- docs: preserve CSA2 boundary case rationale outside test comments
  cwschroeder
  committed
  d30f550
- docs: document reference precision and rounding boundaries
  cwschroeder
  committed
  bed52ce
- docs: clarify selection and retained mHC reference controls
  cwschroeder
  committed
  c7588b1
- docs: preserve execution contracts and satisfy lean comment policy
  cwschroeder
  committed
  a8d4e28
- fix: free failed DeepSeek model allocation exactly once
  cwschroeder
  committed
  c9a31ee
- test: probe TCP exchanges after configurable idle gaps
  cwschroeder
  committed
  eede144
- diagnostics: time DeepSeek TP exchange phases on request
  cwschroeder
  committed
  736236a
- test: exercise DeepSeek framed transport without loading a model
  cwschroeder
  committed
  40cf407
- fix: use the runner clock in the framed transport probe
  cwschroeder
  committed
  b83565e
- diagnostics: separate TP packing and Metal submission waits
  cwschroeder
  committed
  e84be8a
- fix: avoid reserved keyword in TP timing variable
  cwschroeder
  committed
  8530a87
- diagnostics: distinguish GPU scheduling from completion latency
  cwschroeder
  committed
  aa7f09f
- deepseek-v41: use model-owned resources for TP command buffers
  cwschroeder
  committed
  8ee5f9b
- deepseek-v41: request residency after adding TP descriptors
  cwschroeder
  committed
  fe5962a
- deepseek-v41: add opt-in bounded weight residency sets
  cwschroeder
  committed
  b338291
- deepseek-v41: allow periodic residency requests for immutable weight sets
  cwschroeder
  committed
  245ff50
- deepseek-v41: copy only initialized residency handles
  cwschroeder
  committed
  e7abd3b
- deepseek-v41: own residency refresh worker within the model module
  cwschroeder
  committed
  b27e39a
- Interleave stepped GLM CUDA prompt admission with decode
  ashhart
  committed
  b10231b
- docs: record M3 Ultra TCP pair parity, state checks and slower decode
  cwschroeder
  committed
  5aedc7c
- Merge 1.1.1 into main
  ashhart
  committed
  47ea1bb
- app tests: keep the test certificate's serial a minimal DER integer
  ashhart
  committed
  d1018b3
- deepseek-v41: integrate the Metal capability with current shared family owners
  cwschroeder
  committed
  4e20a5e
- deepseek-v41: use Zig 0.17 type reflection in the Metal capability view
  cwschroeder
  committed
  e183f0d
- deepseek-v41: enumerate field names through the pinned compiler API
  cwschroeder
  committed
  0bd7c98
- deepseek-v41: share tokenizer ownership and retain template qualification targets
  cwschroeder
  committed
  f6f7c5a
- deepseek-v41: prototype parallel key scoring without changing attention reductions
  cwschroeder
  committed
  61ab890
- deepseek-v41: prototype whole-slot MoE partition with one exchange per layer
  cwschroeder
  committed
  452ee7c
- deepseek-v41: register parallel attention gates and reserve its 256-thread launch
  cwschroeder
  committed
  cdfa520
- docs: record M3 attention gains and whole-slot TCP partition receipts
  cwschroeder
  committed
  8c6e65a

Commits

Commits on Oct 18, 2026

Commits on Oct 19, 2026