lanes: a family's round planner drives the lane host with prompt pieces, and a halt ends the rounds at a drain deadline (DeepSeek-V4.1 stack 6/14) by jayleaton · Pull Request #662 · ashhart/TensorFold

CUDA keyed graph cache and device arena, both off until a family installs them.

jayleaton

committed

3688f3c

A failed CUDA kernel launch names its kernel and launch configuration; CUresult names fall back to a table when the driver cannot answer.

jayleaton

committed

b7be793

fix(cuda): correct RDMA VMM capability attribute

jrb-ai

authored and

jayleaton

committed

faf5901

Graph cache: a capture is freed until the map holds it, the keep floor stays under the max, a max of 0 is eager

jayleaton

committed

3cc0397

CUDA arena: the address maps are rehashed before tombstones fill them

jayleaton

committed

c58d83c

Tensor parallelism across CUDA processes: bootstrap, NCCL, mailbox and RoCE collectives, plan link, fail-fast.

jayleaton

committed

25c02d3

TP tests: a drain on rank 0 then its clean stop ends rank 1's loop at the stop word with no fail-fast exit; a rank killed mid-drain still fails the other fast.

jayleaton

committed

6b5404d

TP RoCE: with no TF_TP_ROCE_HCA, each rank picks the ports on an IPv4 /24 the peer also has

jayleaton

committed

06c4e75

TP README: when -Dtp-nvcc, TF_TP_ROCE_HCA, NCCL_IB_HCA, NCCL_SOCKET_IFNAME and TF_TP_MASTER are needed, with a two-host example

jayleaton

committed

062af8b

TP test runner: a binary built without the mailbox kernel refuses the kernel backends up front; run_hosts.sh passes HCA picks

jayleaton

committed

d7510ef

TP RoCE: every RDMA post is bounds-checked before the verb, and peers with another region layout are refused

jayleaton

committed

e16e331

TP RoCE: no RDMA write past 1 MiB, and boot refuses any setting that would ask for one

jayleaton

committed

649e788

TP RoCE: TF_TP_ROCE_CHUNKED_KB gathers strides past the slot in slot-sized ops, at most 2 MiB in flight

jayleaton

committed

664a8e2

TP RoCE: a failed transport halts its QPs, drains them and keeps its region; the region is DONTFORK and NOHUGEPAGE

jayleaton

committed

4ad5029

TP RoCE: tf-tp-test failfast V transport checks the failed-transport path on real HCAs, only on an idle link

jayleaton

committed

9bbc21a

Fix RoCE test smaps lookup for merged mappings

jayleaton

committed

cf19042

Sessions: the page map is rehashed before tombstones fill it

jayleaton

committed

0ba7d30

Let a family's round planner drive the lane host with prompt pieces

jayleaton

committed

0918809

Lane host halt: a stop's drain deadline ends the rounds at the next boundary, and every request queued, filling or running fails with the reason.

jayleaton

committed

361f063

Round planner test: both prompts queue before the host starts, and its waits time out

jayleaton

committed

e9f9210

TP RoCE: run_hosts.sh gives NCCL the RoCE picker's ports when neither TF_TP_ROCE_HCA nor NCCL_IB_HCA is set

jayleaton

committed

989ca7e

Session memory for paged-KV families: page pool, prefix trie, RAM residency budget, streamed NVMe tier.

jayleaton

committed

05c5213

Navigation Menu

Repository navigation

lanes: a family's round planner drives the lane host with prompt pieces, and a halt ends the rounds at a drain deadline (DeepSeek-V4.1 stack 6/14) - #662

lanes: a family's round planner drives the lane host with prompt pieces, and a halt ends the rounds at a drain deadline (DeepSeek-V4.1 stack 6/14)#662
jayleaton wants to merge 22 commits into
ashhart:mainfrom
jayleaton:stack/06-lanes

Commits on Oct 11, 2026

CUDA keyed graph cache and device arena, both off until a family installs them.

A failed CUDA kernel launch names its kernel and launch configuration; CUresult names fall back to a table when the driver cannot answer.

fix(cuda): correct RDMA VMM capability attribute

Graph cache: a capture is freed until the map holds it, the keep floor stays under the max, a max of 0 is eager

CUDA arena: the address maps are rehashed before tombstones fill them

Tensor parallelism across CUDA processes: bootstrap, NCCL, mailbox and RoCE collectives, plan link, fail-fast.

TP tests: a drain on rank 0 then its clean stop ends rank 1's loop at the stop word with no fail-fast exit; a rank killed mid-drain still fails the other fast.

TP RoCE: with no TF_TP_ROCE_HCA, each rank picks the ports on an IPv4 /24 the peer also has

TP README: when -Dtp-nvcc, TF_TP_ROCE_HCA, NCCL_IB_HCA, NCCL_SOCKET_IFNAME and TF_TP_MASTER are needed, with a two-host example

TP test runner: a binary built without the mailbox kernel refuses the kernel backends up front; run_hosts.sh passes HCA picks

TP RoCE: every RDMA post is bounds-checked before the verb, and peers with another region layout are refused

TP RoCE: no RDMA write past 1 MiB, and boot refuses any setting that would ask for one

TP RoCE: TF_TP_ROCE_CHUNKED_KB gathers strides past the slot in slot-sized ops, at most 2 MiB in flight

TP RoCE: a failed transport halts its QPs, drains them and keeps its region; the region is DONTFORK and NOHUGEPAGE

TP RoCE: tf-tp-test failfast V transport checks the failed-transport path on real HCAs, only on an idle link

Fix RoCE test smaps lookup for merged mappings

Sessions: the page map is rehashed before tombstones fill it

Let a family's round planner drive the lane host with prompt pieces

Lane host halt: a stop's drain deadline ends the rounds at the next boundary, and every request queued, filling or running fails with the reason.

Round planner test: both prompts queue before the host starts, and its waits time out

TP RoCE: run_hosts.sh gives NCCL the RoCE picker's ports when neither TF_TP_ROCE_HCA nor NCCL_IB_HCA is set

Session memory for paged-KV families: page pool, prefix trie, RAM residency budget, streamed NVMe tier.