LunNova

Thank you to tiny corp for pointing out some problems running BERT training with Tinygrad on AMD GPUs in this Tweet. We had a few engineers at AMD take a look at the problem and they were quickly able to reproduce it.

What they found was an issue related to CWSR (compute wave save restore), which is a mechanism that allows our driver and firmware to preempt and reschedule long-running compute waves on our GPUs. The GFXv11 GPU line requires a workaround to set COMPUTE_PGM_RSRC1.PRIV=1 when dispatching a compute kernel. Normally this is handled by the AQL DISPATCH packet. However, since the Tinygrad implementation leverages a custom runtime, it requires this workaround in its PM4-based dispatch. This patch is specific to GFXv11 GPUs. Other GPUs do not require it and should not use this workaround. The following KFDTest patch can be used as a reference: https://github.com/ROCm/ROCT-Thunk-Interface/commit/507637ed5b82197eecbf483cdc1234939766549a

While inv

LunNova

/// Guards a scope against unwinding, calling a handler if unwinding occurs.
///
/// - Handler gets no panic info; limited to `Fn()`
/// - Not reentrant; handler panics may cause program abort
/// - Intended for single scope; don't store or share
/// - Ineffective with panic=abort
pub struct UnwindDetector<T: Fn()> {
    handler: T,
}

LunNova

function fp_to_bytes(fp, bytes, is_double)
  local val = tonumber(fp)

  -- it's a NaN or inf
  if val ~= val or val == math.huge or val == -math.huge then
    bytes[1] = (val ~= val or val == math.huge) and 0x7f or 0xff
    bytes[2] = (val ~= val and 0xf9 or 0xf8)
    local max = is_double and 8 or 4
    for i = 3, max do
      bytes[i] = 0

LunNova

Run in MappingTest subdirectory of https://github.com/MinimallyCorrect/Mapping/tree/transform-only-runs-for-default-attr
Must run publishToMavenLocal in parent directory first.

Note how transform only runs in test case -PTEST=3 where the default value of the attr can be transformed to the requested value.
Even case 4 fails which is where the default value is set as in case 3, but a value the same as that default value is also set in the gradle module metadata.

$ for i in 0 1 2 3 4; do echo; echo "Testing with -PTEST=$i"; echo; ./gradlew.bat -PTEST=$i --no-build-cache build; done

Testing with -PTEST=0

LunNova

// ==UserScript==
// @name PluralSight Wider Speed Range
// @namespace https://nyx.nova.fail/
// @version 1.1
// @description Faster pluralsight video maximum speed
// @author Luna
// @match https://app.pluralsight.com/*
// @grant none
// @run-at document-start
// ==/UserScript==

LunNova

#!/usr/bin/env bash
# This script boots the first linux it finds in a grub config on any available mount point
# using kexec

DEBIAN_FRONTEND=noninteractive sudo apt-get -q -y install kexec-tools

set -euvo pipefail

parts=$(sudo blkid | grep -v /dev/loop | sort | cut -d: -f 1)
grub_paths="/boot/grub/grub.cfg /grub/grub.cfg /boot/grub.cfg"

LunNova

Enclosure serial console commands

COMMANDS:

Note: Not all commands are supported, 'help' lists supported commands.

acfail_sim Simulate AC Fail condition
auto_iic_recovery Enable/disable the automatic bus recovery.
batt_cell_balance Override auto cell balancing by individually turning on/off the balance FET s
batt_clear_cell_fault_poh Clear the Power On Hours stored as a result of a >500mV cell imbalance fault

LunNova

Jul 11 09:22:21 	dhclient[87943]: exiting.
Jul 11 09:22:21 	dhclient[87943]: connection closed
Jul 11 09:22:13 	dhclient[27206]: em2 link state up -> down
Jul 11 09:21:26 	dhclient[73751]: bound: renewal in 190763 seconds.
Jul 11 09:21:26 	dhclient: Deleting old routes
Jul 11 09:21:26 	dhclient: Comparing Routers: Old: <FLAPPING GATEWAY IP> New: <FLAPPING GATEWAY IP>
Jul 11 09:21:26 	dhclient: Comparing IPs: Old: <FLAPPING WAN IP> New: <FLAPPING WAN IP>
Jul 11 09:21:26 	dhclient: Starting delete_old_states()
Jul 11 09:21:25 	dhclient: New Routers (em2): <FLAPPING GATEWAY IP>
Jul 11 09:21:24 	dhclient: New Routers (em2): <FLAPPING GATEWAY IP>

LunNova

# strace -x -s 256 -p 1957 -p 1958
Process 1957 attached - interrupt to quit
Process 1958 attached - interrupt to quit
[pid  1958] rt_sigsuspend([] <unfinished ...>
[pid  1957] rt_sigsuspend([] <unfinished ...>
[pid  1958] <... rt_sigsuspend resumed> ) = ? ERESTARTNOHAND (To be restarted)
[pid  1958] --- SIGRTMIN (Unknown signal 32) @ 0 (0) ---
[pid  1958] sigreturn()                 = ? (mask now [])
[pid  1958] rt_sigprocmask(SIG_BLOCK, [CHLD], [RTMIN], 8) = 0
[pid  1958] rt_sigaction(SIGCHLD, NULL, {SIG_DFL}, 8) = 0

LunNova

rdrand stressor will be skipped, not a recognised Intel CPU.
tsc stressor will be skipped, not a recognised Intel CPU.
disabled 'cpu-online' as it may hang the machine (enable it with the --pathological option)
dispatching hogs: 16 af-alg, 16 atomic, 16 branch, 16 bsearch, 16 cache, 16 context, 16 cpu, 16 crypt, 16 fp-error, 16 funccall, 16 getrandom, 16 heapsort, 16 hsearch, 16 icache, 16 ioport, 16 lockbus, 16 longjmp, 16 lsearch, 16 malloc, 16 matrix, 16 membarrier, 16 memcpy, 16 mergesort, 16 nop, 16 numa, 16 opcode, 16 qsort, 16 radixsort, 16 str, 16 stream, 16 tree, 16 tsearch, 16 vecmath, 16 wcs, 16 zlib
stress-ng-numa: system has 1 of a maximum 1024 memory NUMA nodes
stress-ng-stream: stressor loosely based on a variant of the STREAM benchmark code
stress-ng-stream: do NOT submit any of these results to the STREAM benchmark results
stress-ng-stream: Using CPU cache size of 8192K