Commit graph

12,415 commits

Author SHA1 Message Date
George Hotz
6eee1a161b
Merge branch 'master' into sym_work 2026-02-26 16:20:50 +08:00
George Hotz
7bc9ebf201 real tests 2026-02-26 16:20:15 +08:00
George Hotz
1c8517d1a3 fix after in the big graph 2026-02-26 16:14:59 +08:00
qazal
12d179f5f4
viz: brighter call.src[0] edge color (#15022)
* work

* 2

* better color
2026-02-26 16:07:22 +09:00
George Hotz
5a6790e58b fix symbolic shapes in calls 2026-02-26 14:41:21 +08:00
George Hotz
2655655a0c
call gradient creates a call (#15020)
* function creates a full subgraph

* tests

* fix var

* fix tests

* implict assign/contig

* move kv init
2026-02-26 14:15:29 +08:00
Christopher Milan
94acd85285
fix typo in nn/__init__.py (#15019) 2026-02-25 20:01:32 -05:00
Christopher Milan
e5c0db66d1
num_batches_tracked does not need is_dtype_supported (#15018) 2026-02-25 19:50:57 -05:00
George Hotz
3244131f59
update dagre with more recursion fixes (#15012) 2026-02-26 08:35:05 +08:00
chenyu
ed9d475a12
assign tests with test_function (#15015) 2026-02-25 16:15:59 -05:00
nimlgen
faa66e0a61
mi350 hive_reset am repro (#15014) 2026-02-25 21:30:18 +03:00
nimlgen
8983830aa8
am: code style consistency (#15013) 2026-02-25 21:30:10 +03:00
George Hotz
0d35b67f2c
revert realize to only be buffers (#15008)
* revert realize to only be buffers

* fix that

* broken attention

* Revert "broken attention"

This reverts commit a23c3cd96c.

* and that
2026-02-25 22:43:06 +08:00
qazal
35f85c393f
viz: keep recursively nested call collapsed (#15010) 2026-02-25 22:45:18 +09:00
qazal
421b1d4a56
viz: monospace font for tags, no dy overrides (#15009)
* viz: monospace font for tags, no dy overrides

* str
2026-02-25 22:15:31 +09:00
qazal
448e997be4
gemm/asm: cleanup custom function args (#15007) 2026-02-25 22:05:56 +09:00
qazal
c58e91942c
viz: support collapsing individual CALL nodes (#15006)
* all

* contracted all by default

* simple call mask

* work

* minus not hyphen

* color / cleanup

* detail
2026-02-25 21:27:25 +09:00
George Hotz
68831cd852
add more tests to test_function (#15003)
* add more tests to test_function

* add function to llm

* function decorator on llm

* works

* symbolic fixups

* minimum change

* implicit inputs

* don't actually update llama yet
2026-02-25 18:42:06 +08:00
wozeparrot
d941dd5aeb
llama3: pad vocab when mp sharding (#14998) 2026-02-25 00:04:06 -08:00
wozeparrot
e1c9985715
llama3: better time keeping (#14999) 2026-02-24 22:42:05 -08:00
Christopher Milan
4a2fc7ecbb
autogen: cache downloads (#14997) 2026-02-25 01:34:27 -05:00
George Hotz
e3fa9896b7
start function and add walk rewrite (#14992)
* start function and add walk rewrite

* work

* add function on feed_forward

* llm progress

* stuff

* none of that
2026-02-25 13:56:27 +08:00
chenyu
fde7a40bb0
allow dtype mismatched assign on disk (#14993)
reverted #14473, that was a bad idea. also added a test that safe_save only has copy
2026-02-24 20:49:55 -05:00
chenyu
46d9a9a74f
minor indexing cleanups [pr] (#14991) 2026-02-24 16:49:35 -05:00
chenyu
8dae9be573
move realize_map fixup into realize_assign_src [pr] (#14990) 2026-02-24 15:51:40 -05:00
chenyu
9d9151a21e
remove const normalization in indexing [pr] (#14989)
rangeify can create const with device, and all is normalized in to_define_global
2026-02-24 15:09:11 -05:00
chenyu
f68a472244
end range for COPY/BUFFER_VIEW [pr] (#14987) 2026-02-24 13:33:35 -05:00
chenyu
e5d27a3773
remove BUFFER_VIEW from ended_ranges special case [pr] (#14986)
* remove BUFFER_VIEW from ended_ranges special case [pr]

* will fix later
2026-02-24 10:37:29 -05:00
chenyu
5fd4fc0c6d
fix tinyfs (#14974)
* fix tinyfs

* fix that
2026-02-24 08:50:53 -05:00
George Hotz
8a6dffc87e
Tensor.callify will be the JIT (#14983)
* close

* simple callify, support linear in the scheduler

* all tests pass

* everyone is happy

* dumb test

* Remove unnecessary blank line in rangeify.py
2026-02-24 18:42:24 +08:00
nimlgen
6f1cb6be86
am: tiny err handling cleanups (#14981)
* am: tiny err handling cleanups

* x

* x
2026-02-24 12:43:45 +03:00
George Hotz
b643fca51e
clean up complete_create_schedule_with_vars (#14980)
* clean up complete_create_schedule_with_vars

* transform_to_call

* update viz tests
2026-02-24 16:12:36 +08:00
wozeparrot
8d9545e09e
llama3: correctly shard wqkv (#14978) 2026-02-23 23:57:10 -08:00
wozeparrot
a36a26d4ed
llama3: optim does grad acc in correct order (#14965) 2026-02-23 22:25:13 -08:00
George Hotz
e2b1f2620d
schedule is linear (#14975)
* schedule is linear

* cleanup

* cleanups
2026-02-24 11:30:41 +08:00
Christopher Milan
57ade7608a
consider indexing math cost for IMAGE=1 (#14973) 2026-02-23 18:57:45 -05:00
chenyu
0bda5585c7
unit test TestTinyFS (#14972)
these passed before the allocation change
2026-02-23 16:59:39 -05:00
imaolo
405d37423e
call release() in MetalAllocator._free (#14970)
* add failing test

* call MTLBuffer.release() in MetalAllocator._free()

* Update test_metal.py

---------

Co-authored-by: nimlgen <138685161+nimlgen@users.noreply.github.com>
2026-02-23 23:33:31 +03:00
nimlgen
77db8e1c07
cpu: wait on dep signals (#14862)
* cpu: task_done() in case of failures

* print

* fix

* x

* f

* x

* um

* ?

* u

* f

* x

* gh

* f

* f

* virt

* x

* simpler
2026-02-23 21:09:41 +03:00
chenyu
127136421d
enable a few WEBGPU isnan tests that work now (#14967)
* enable a few WEBGPU isnan tests that work now

* still failed
2026-02-23 11:06:08 -05:00
ttomsa
0366474089
Bool cast to cmpne (#14544)
* test

* rm in llvmir

* rm in ptx and nir

* hmmmm

* rm in decompositions

* skip tests

* add test

* just this

* rm comment

---------

Co-authored-by: chenyu <chenyu@fastmail.com>
2026-02-23 10:31:36 -05:00
George Hotz
806581f807
rename rewrites + sink filter + bump to dagre 2.0.0 (#14966)
* bump to dagre 2.0.0

* transform to call

* cleanup names

* get kernel graph

* dagre recursion fix + better error

* add toggle to hide sink nodes

* no sink by default

* revert that

* only hide final sinks

* lol
2026-02-23 22:47:22 +08:00
nimlgen
d86f1d66b5
system: apl validate dev_id bounds (#14964) 2026-02-23 12:18:03 +03:00
George Hotz
b824490e3f
allocate generates a call (#14958)
* allocate generates a call

* symbolic works too

* DEFINE_VAR is param

* replace param later

* apply buffers

* name

* upd

* this was a bug...
2026-02-23 15:59:20 +08:00
wozeparrot
dd8302a6d0
fix: optim device is never none here (#14963) 2026-02-22 23:34:57 -08:00
wozeparrot
25565b2410
fa: test for mp (#14907) 2026-02-22 21:47:36 -08:00
qazal
d6145736c7
sqtt: examples generator changes from inst_discovery (#14961)
* sqtt examples generator changes from inst_discovery

* rdna4

* rdna3

* cdna

* sad reality for mi300x
2026-02-23 14:42:48 +09:00
George Hotz
3acd763684
simple call in allocate (#14962)
* allocate generates a call

* symbolic works too

* add min/max to PARAM

* revert viz
2026-02-23 13:34:20 +08:00
George Hotz
f45199269b hotfix: regress NV cifar_10steps_half to 120 ms 2026-02-23 12:29:25 +08:00
George Hotz
677145b393
all consts have shapes (#14959)
* all consts have shapes

* vconst has shape too

* use normal schedule

* cast ptrdtype

* image

* bitcast issue + hack
2026-02-23 10:26:50 +08:00