Retrospective DCO confirmation by Martin Vit for 131 commits

The issue text contains the retrospective DCO confirmation and the list of 131 covered commits. Read the repository’s contribution or DCO guidance to determine how retrospective confirmations are recorded; the issue names no file or test to start from. Done means the confirmation is handled for the listed commits through revision e774a468cee4ade64393ba9d856788aaef974ed7, without extending it to future commits.

I have reviewed the Developer Certificate of Origin 1.1 and certify that my contributions in each of the commits listed below satisfy its terms under the applicable open-source licenses.

Repository: https://github.com/NVIDIA/hyperflare-inference

This retrospective confirmation covers the 131 commits explicitly listed below, attributed in Git to Martin Vit with email [email protected], through main revision e774a468cee4ade64393ba9d856788aaef974ed7. Commits that already carry my DCO sign-off are excluded. This confirmation does not cover future commits.

Signed-off-by: Martin Vit [email protected]

Covered commits

1a61ce3b1ba84cf4a703b2c06be34386624e626a  Write speculative completion records before mutating request state
b5355fed6aa5da4dc3436c9cddef6a7762d7ce15  Give MTP drafts an NVFP4 copy of the target lm_head
f5e3b1314f9e87dad99b9d19a7fc5876b281c941  Serve the loaded model for unknown, empty or missing model names
4f5ce3f1cd8d5d85a472911bf7bb55eb17bb1d8b  POC: choose speculative width from measured per-depth acceptance
ddc1afed7d4b53e91c60e13ec04151a68737d7e5  Draw Qwen MTP drafts from at most 64 top-k candidates
a5e118be42c5b3628194c465d0045496a0e51112  Pack Qwen MTP draft metadata in one kernel
bef30c52ddddded3a31bc95477b02400265ba485  Give GLM MTP drafts an NVFP4 copy of the vocabulary shard
32c4374afd934b32182ace9b59613bb6b470e34f  Write GLM KDA decode output directly into projection storage
aea86a09644e6a0a1915de4bddcab8f034f8261d  Commit accepted GLM KDA states in one launch
6fce572fd3e7916748ddb44182add4d2f6eaec40  Draw GLM MTP drafts from at most 64 top-k candidates
82fe86b287c38679891cfd07fb550833796a8149  Size GLM KV pages from memory measured after graph capture
aa7a19e82a6dc697fa24e22e239a260b443a1107  Size compiler workers by free device memory
1dbbc4dfabe6d386c9b473d17422c7840d8c85ae  Share GLM draft-layer scratch across serial passes
5bf4c469e2fa6cbe122dccccdbc187326641943d  Convert GLM mHC parameters to FP32 once at load
92e2855446955cbf000feb575706a3dbb0efdf53  Let the GLM model contract set global tokens per KV page
d94499743eecf967c28c543f2d0c395557735bd6  Shard GLM sparse-attention context across tensor-parallel ranks
4552319e857d2b77fd983a909fb30ecd02948aa2  Release CUDA contexts that checkpoint reads open on peer GPUs
a39fde15c613d29bea43b248737f8b9434440c6b  Keep GLM speculative state rows for the widest verification only
3708dd96afe95b77a04c51a427ec9b3c0ab816ea  Select GLM prefill index pools in 512-row slices
c9ed9cb08bfa340e591acce69e5d6da1f74619a4  Absorb, attend and expand GLM context-parallel queries per slice
3d2df3b1e728e4455abf42896a29e67f5d11da75  Overlay GLM prefill layer scratch and keep decode scratch separate
6c52d1b1816d73b4b643aec2a2d2dfdf95ca6d2b  Test GLM context-parallel placement and selection kernels on CUDA
03fc6984cb00c032f85cbe7ac74f42bc225712d2  Prefill GLM context-parallel attention over gathered context blocks
cb673c45f94d4e1b680d50105e64cdf107a46a5b  Export GLM context-parallel packages through package storage
6ecf347a272b3864d1068d86e20ac0cf395cd958  Serve any prepared model contract from flare run
e6d19c896a373ef79f645fd2c893f45dc8dc2b78  Export graph copies that write another rank's peer storage
810fa38d854740e8086f924359f3b21d83ae675c  Return transient preparation memory from package storage
c848bd6c7afbf18c01aa2301fc6cf8cd7e5980b5  Attend GLM context-parallel prefill over the whole gathered context
e5259271518b99b692c23dc27fa16ca66722514a  Fuse the GLM KDA gated output norm in prefill
dcb1d7b1e18f5e0d5c286f2ef0405bcd744dfe00  Keep an unset shared-memory carveout unset in exported graphs
0277b52daab6414d64685a2f92b55bac65051b18  Return Qwen's transient preparation memory before sizing KV pages
b8393309472cc197106e026e992b9a7640f95eca  Prepare prefill graphs in steps so partial chunks pad less
a32c650b3a9e6486be530e64c1f7f6ba258f08ee  All-reduce GLM prefill chunks through the copy-engine ring again
f5b539b03833d74937224c50d087577aa1cd1199  Hand the package allocator to the copy-engine ring directly
7de7eb35373f6ba7e834529a3edb1afc9edbe279  Size the request body limit from the model's context
8ffe32514ba3e55b514ee66fb862a375e60d0aef  Let exports keep attention context whole on every rank
2b5a4b39640623c1e3efb84750540379f21232e5  Merge GLM context-parallel candidates by one bitonic step
d37939a7a731b6f62fb505395d62d6460edd6648  Project and gather GLM decode queries beside record selection
890404ec16fa623a53202b1eaa02d6be79ba22ab  Read GLM sparse attention outputs in place
9d7aafc6a04f03b2677c5514f3b3eb4af6465300  Fuse GLM index key and head-weight projections into the layer input GEMM
6c9d0dcda7cea0cc0105b91b9070b5aa9c9489df  Fill GLM index tables once per forward
8547b933e6be7621c92d04142511bff666f6a3ad  Score GLM draft proposals in two fused kernels
54a78dfa0cbf8295097e75d3560de524a75edd8f  Place GLM draft rows in one kernel
f86d520c83085b1b9cc5f04fe61f61fb60938d62  Optimize DeepSeek V4 TP2 graph execution and prefill storage
4adf270d1ae6aeb825c517356716f1089d2c3a10  Document DeepSeek V4 tensor parallelism and cache geometry
cdc02d1a57cf43cc3724a34930fc860db348e498  Prepare forwards on explicit compiler GPUs and serve diagnostic DeepSeek graphs
97dca6b64e610b4b9cd972aea033080d25467273  Bound metadata compiler workers independently of guarded intra-op settings
8c7d817be5f68b5e9ecbdbf7f0df57287370b80e  Integrate structured-answer support with DeepSeek V4 preparation
0446be7d9abda7d6f9dfbbeac81c06cfe333b0c1  Bound tuning fixture lifetimes by live candidate calls
81ea12e61f89dc61c3ad93e3bdb8a84b975ecb3b  Support shared RAM embedding tables in development serving
f33eec48139bd5c05500d4500ec22f077599cfec  Grow GLM pages on demand and preempt the newest request when they run out
ad9b2d30aeb235b09df66075848555316a1e5a39  Grow pages for every GLM text request, whatever its output limit
24cd5f9a83221e4e3756c8e54f488572f876ef11  Sample a preempted request's continuation from a seed derived at preemption
355c7a82d7e8666df674da7b60ddd1d46349f891  Leave GPU memory for the host prefix store's registrations
f2d959e390da6148476cb31ca6da53d750544830  Split a memory-only host prefix store by rank
6b2d25f7d58c1351dee01e05062605fdb4a4e47c  Prefault resident embedding table mappings before row delivery
edcb8be3c50313b049fd80c49724871688b207a2  Batch DeepSeek V4.1 sparse selection over 128 query rows
1054e935604c83076dd574fa3d692b33796c4e94  Batch accepted-prefix copies across continuation rings
4e36eb3907e64327d19841b3b738bfbf70eeb531  Preserve higher draft scores when cutoff ties fill the candidate buffer
4f0c3eea9dcd5306e41b87508787cf86c8ffee5b  Treat signed zeros as equal draft scores
f43578d2f4a2fbfade1d8adde8eb0194be41e7e0  Bound DeepSeek V4.1 decoder replay and batch accepted cache updates
a5a97e0269b5d43486022097576e766332dae730  Integrate typed prepared contracts with DeepSeek execution optimizations
af05d1c662016a08483318fd13cdec767a80240a  Isolate FlashInfer JIT workspaces by installation
63f45f3a74dfeaca38bcb967e27514992a0b9dda  Merge native Windows support with shared resident embedding tables
24bee224b723e54b4dc6b259e69500a12a580613  Integrate Linux package address isolation with DS4.1 preparation
caac00aeb74e070804e2cd0b89204b1734dfa4c7  Keep DS4.1 CSF scales resident with shared graph storage
ff32cecd310b4a115a51a81e53b195445862e33d  Load DS4 Flash CSF through shared MXFP4 expert storage
56741de1fc825fc41c61fcf35f543668208928a2  Add opt-in causal split speculation for native single-request research
4925d10aeb19898904027f8b14667471bb4da50d  Fuse Kimi K3 residual mixing and routing; load native MXFP4 experts
78c1d491f392d1a56cc632c732ecfc0afbbd6ca9  Assemble exported packages from staged rank files
84c1c0a1195810a5c251c6284f18eccb616b0f5b  Lend missing forward compilations to idle ranks on one host
fa300a28e8304ff3dcfc5b6bbd19a1f288ea679e  Connect single-host rank groups over loopback
b9d767a2cb4c35c0e5b8985e757198399d1975ad  Read checkpoints through bounce buffers without nvidia-fs
418e2e4e2829e58f8fccf1a72a41ad3259149d92  Read column-sharded checkpoint tensors once per host
91bdb696c2e099f1125e3a60fa6d705604babcf6  Exchange E4M3 head shards and propose greedy drafts over vocabulary shards
38685195aac0599ea4b1b4a58297885210f6fd4a  Serve Kimi K3 with decode context parallelism, E4M3 latent KV and DSpark
894623087e99e63838f85b7ff5658e6c5eb8f714  Shard Kimi K3 MLA latent projections across ranks
8c225aafe69260291f6130854cbc102053442b84  Gather Kimi K3 prefill context one window at a time
3485a754c959fd47f9d274578b796d2d5872b1de  Add --language-model-only for recipes with a separable vision encoder
f56a2a1c7712deca75ef63b62892af07fb7e4094  Declare measured Kimi K3 serving memory
b3bd121efc29610b6cf0a6253df9a427a5f3792e  Reduce Kimi K3 prefill activations in chunks of a decode-sized channel
b9b06dde6c47fe35a0af86a77439379939cf8ef1  Gather Kimi K3 prefill activations in chunks and share attention outputs
b0274b13c226571bb2e88a37df9713756b9d4161  Back only the live granules of exported preparation allocations
28730f2da7dfe2a8dbd7574c810b59e3fabe5851  Reuse consumed Kimi K3 activation buffers for mixtures and expert outputs
3d19a15781d29bedc75dab50287f9b164e64c8d0  Serve Kimi K3 DFlash2 beside a context-parallel target
9a7196118821cabe601b44ef2afd26b0de77c426  Declare Kimi K3's b12x settings in its recipe
b5409b559d1f648241eee5f39416f12e9848e5e9  Use root proposal confidence without peer numeric consensus
57fd128abb0f23c62cda0204b5f93729fced54a3  Age proposal forecasts toward population confidence and widen cohorts
a2b9c41ceeb35c2975258cd33417f4122dd4f6c1  Align DeepSeek V4.1 prompts and cache lookup with checkpoint 1017
3502d37b972210c1be7036f31877df07aa47fbd7  Respect RAM checkpoint storage and bounded diagnostic preparation
32ace57fcd574e03210e0a9655ad218de18123e5  Prepare DeepSeek V4.1 image graphs with bounded feature slots
293643b435e2a87eac4724c82f3f62a068d9ffa2  Declare DeepSeek V4.1 image admission capacity
85bc2de7e95b3272a89bdb0f133d01841fd5e1cf  Reduce DS4.1 speculative attention and projection work
1f4da7a55cb90dd732dfaab299ce8dce9be7febc  Select retained proposal depths from measured execution costs
f24008706c01a1075980d70ed904c1594b35bd08  Poll brief tensor-parallel command handoffs before parking
9433368863b7b1954d926168ce04b92c455f9500  Collect compiled-forward loading temporaries at phase boundaries
c8e4715ab7fb23be7a83b6e283d451f75e8a1617  Read publishing licenses from checkpoint metadata directories
1100b83c29ec23a5480c159ddd7498e538a2da93  Skip inactive completion payloads while polling GPU work
8dca90a5e7c94e611b2a36638ee33c951bc3d17a  Support Triton 3.8 kernel argument metadata during export
95a97aeaf6792ae1fbce16bc108198337d9ee38c  Add packed verification with frozen scheduler row budgets
6ad1582b633bf340912bf8bfdc2b95c774da36c3  Report cumulative native token usage in completion streams
cf9c3ddd66aed6001c82288f008c7937da91e7b5  Reduce native graph input and proposal-feedback host work
022c33bfe07d1395dc377608f0e6a987bb591648  Project packed vocabulary rows before expanding sampling slots
24fd9f1c12b7f34b8c8fd87b7bf03ad483a7540c  Fuse typed result broadcasts and size a sampling channel from payloads
ea0a91b8ec8b15f575f5ecfd0b63be0419ffd4a0  Expose BF16 MLA preparation constraints with a pinned FlashInfer API
cb786267b7fab10484f73b7333fef0c551a161f9  Retry transient busy seals before publishing resident embedding tables
c5f920cc6c729164fcb6343c3f42d8aa20968bbf  Reject unavailable diagnostic serving and invalid export metadata early
ae8ae46b395c49d3e61bc741518190403c4c4723  Reduce prepared decode memory traffic and disabled-filter work
83d582e2925c103a92452f27bffaf817e2362229  Keep parallel tool-result images aligned with their calls
bcba6b087dd0beb6b82e978f2afdd7ef9e82a7d9  Publish tested multiarch serving containers from main
3e102fe6b840fdced18b1510a0623984ce60e76c  Announce published Flare changes with a durable six-hour interval
bf5333ffe4e75740028bcc22eba5a963330aa141  Check the orbit command in container smoke tests
2d0c3f41f09426c00e7de9602f8fb5b24dfbc646  Constrain changelog generation to validated prose openings
1ec86b54d3bd9573f6d5df120a74502128320716  Support numeric reasoning effort for DeepSeek V4.1 (#50)
328ddce45b38fb0df55272d27ec1032f2d963a8c  test: check sectioned unverified payload inspection output (#89)
9e19b3bac8972f0e8559fac64d266c84c5b1dfe9  Pin FlashInfer to the supported tuning-tier API (#85)
63a4008973fe1934637a046e0ae3c93a9e40ab82  Fuse DSpark draft scoring and sparse support restoration (#95)
d592f1fc70f51234175994a38da92172ad2648b1  Overlap DeepSeek V4.1 query expansion with cache preparation (#97)
9e35eeb22879b6fde39f6cc89ef76d6ab3a9838c  Prepare an opt-in MXFP8 DeepSeek V4.1 draft vocabulary head (#98)
aa885bdf5c3cac37a2bc159856ca2808a802bdbc  Preserve programmatic dependencies in native CUDA graph loading (#104)
2436d9e473ecc114c1757de3c2a034683de6e67d  Broadcast speculative proposals by block and fuse BF16 Markov sampling (#107)
f706f2bac3c09d19096aa1c69c95ae8c65fbe3d4  Prepare DeepSeek V4.1 MXFP8 attention with public b12x plans (#109)
e60bb3bc87fd32238cfcc488499583cfe5f8e82b  Tune DeepSeek attention projections together in complete layers (#110)
98f816827753816198ba45105de9e6022fcad03e  Include FlashInfer CUDA sources in preparation compatibility identities (#120)
56f5702a0700add9fd98771d582effa68d2cf998  Prepare row-scaled FP8 weights for the DSpark transition head (#121)
5d546f13f23880c657900769c4d19c527bc3c40b  Sample and verify distributed vocabulary probabilities (#135)
921afd5cd38bd272a6021b310b44e8e3b27bbb11  Append DeepSeek V4.1 window snapshots during cache writes (#108)
18b16e170bb9eab03e2b1800107c72966c56594f  Prepare clustered BF16 PCIe all-reduce ladders (#174)
78b18b1c378fbd3671c1996cc930f3f0bf93f240  Share startup race storage and reuse resident weights for L2 eviction (#197)
f56cd2bfd595033646f1ff4bafe453319d3af803  Collect forward binding temporaries without repeated full-heap scans
905db6a28067e6da80623d925695f4c4fd178057  Benchmark verification prefixes with complete proposal work