* fix Muon weight decay being a no-op
LARS._step computed the post-momentum weight decayed param but only
used it for a dtype cast, so the decay was never applied. Fold the
decay into the update instead, matching torch's param.mul_(1 - lr*wd).
test_muon_wd passed anyway since lr*wd=1e-5 is far below atol, so
also bump the test's weight_decay to 10 to actually exercise it.
* muon: apply weight decay after lr scaling
keeps the decoupled decay independent of the LARS trust ratio r,
matching torch's param.mul_(1 - lr*wd). no behavior change today since
r is always 1.0 on the pre_wd=False path (Muon has tcoef=0).
* mergable fast RDNA3 Qwen 3.6
* AMD
* quant 256 multiple
* cleanup cast
* llm kernels: adapt to Ops.BIND removal
Variables are 0-d ALU BUFFERs in the tensor graph and take the ALU PARAM form
inside kernels (UOp.variable(param=True)). Add kernel_var helper for the
conversion, and keep start_pos in bound form at the graph level so function
implicit-input collection and the schedule's binds rename-back line up.
* adaptive prefill chunk sizes for recurrent models + iq4xs model entry
one TinyJit per static prefill chunk size: capture 128 and 32 at warmup,
generate picks the largest that fits the remaining prompt. long prompts
prefill 2x faster (555 tok/s on Qwen3.6-27B IQ4_XS) without pushing short
prompts through token-by-token decode.
* minimize diff: early-return custom attention path, keep master state init
* minimize: master _attention with gated fused-scan swap, kernels/amd only, single chunk size
- GatedDeltaNetBlock._attention keeps master's symbolic-padding structure;
the recurrent scan is swapped for the fused gated_delta_prefill kernel only
on RDNA3 with static shapes (fast_scan), everything else uses the old path
- all AMD kernel code lives in tinygrad/llm/kernels/amd.py (drop kernels/__init__.py,
drop the generic fallback kernel - the old scan covers non-RDNA3)
- single prefill chunk size 32; non-RDNA3 recurrent keeps master's chunk_size=1
- the conv+normalize miscompile doesn't trigger with master's window-buffer conv,
so the contiguous workaround is dropped
* warmup: single code path for fast and old recurrent
* drop fast_scan/fast_recurrent flags, inline the RDNA3 gate (cached)
* generate: chunk size is always 32, no device gating
static chunks for recurrent models everywhere: the fused kernel path on RDNA3,
the old scan elsewhere (which is also faster chunked than token-by-token)
* warmup: drop redundant _init_state loop (lazy init in the eager step covers it)
* symbolic-length prefill with the custom kernels
the prefill path is fully symbolic again (master's generate, one prefill graph
for every chunk size, no static-tail decode): padded steps are exact no-ops in
the scan (beta=0, alpha=exp(0)=1), flash attention positions queries at
start_pos instead of valid_kv_len-M, and quant linears pad to the chunk bucket
prefill 401 tok: 284 -> 348 tok/s on Qwen3.8-27B IQ4_XS (tail chunks no longer
decode token-by-token), decode unchanged at 45 tok/s
* cli: default qwen3.6:27b to the fast IQ4_XS quant, add qwen3.8:27b
Q4_K_M falls back to slow inline dequant with the custom kernels, IQ4_XS is
the fast path. qwen3.8 quants use unsloth's UD (dynamic) naming
* warmup: back to master's two-liner plus a cache reset
with symbolic prefill, generate([0])'s 1-token chunk captures the symbolic
prefill graph that serves every chunk size, and JIT batching on capture
measurably doesn't matter with the fused kernels (347.7 tok/s either way)
* cli: pin qwen3.8:27b to the pre-UD revision
the UD-IQ4_XS replacement mixes in Q3_K tensors (ggml type 11) the loader
doesn't support; the pinned revision is byte-identical to the known-good file
* warmup: identical to master
the leftover cache is self-consistent: get_start_pos only reuses a full
strict-prefix match, everything else restarts with a state reset
* model: hoist the quantized_attention import to the top level
* hoist the GDN query scale out of the branch, restore master dtype.py
the scale is the same op in both paths, apply it once after the transpose.
the dtype.py diff was a stale pre-SPEC=2 copy, not intentional work
* gated_delta_prefill: don't pass the bound start_pos as a call src
device-less param buffers in call srcs crash hcq2's _get_enqueue_devs. the
var already reaches the graph through the state AFTER chain (conv state
store), same as the flash kernels' valid_end
* llm: half KV cache with custom flash kernels, drop the int8 quantized cache
matches master's new half cache default: no scales, no packing, one less
buffer. the store casts to half explicitly (buffer-only half usage misses
the renderer's half define). 45.5 tok/s decode, 348.7 tok/s prefill —
same as int8
* llm: zero-init the KV cache
the int8 path was accidentally protected from uninitialized memory by its
zero-initialized scale buffer; with a plain half cache the flash prefill
kernel's P*V wmma computes 0*NaN=NaN on masked lanes past the valid region
(manifested as garbage tokens at 32k context where the allocator reuses
dirty VRAM)
* gate that
* llm/kernels/amd: reorganize by kernel family, drop the clutter
sections: shared helpers, quant linear, flash attention, gated delta prefill.
no AxisType.WEAK (default), no ALLOW_DEVICE_USAGE override (unneeded), magic
numbers become names (QUANT_SIZES, Q5_K/Q6_K/IQ4_XS), merged wrapper layers
(flash_attention_causal_cached folded into flash_attention), one _unbind
helper for the bound-var dance
* test: universal recurrent reuse assertion, fix lambda lint
* 1-token chunks have a static shape: they are decode steps
a 1-token chunk routes to the decode graph via the existing dispatch, so
warmup and decode-only workloads never build the big symbolic prefill graph:
CI benchmark command 12m50s -> 5m29s (master: 6m48s), 220 -> 123 compile jobs
also restores the ALLOW_DEVICE_USAGE override in amd_custom_kernels_supported:
Device[] asserts inside @function contexts (ALLOW_DEVICE_USAGE=0), and the
first gate call can happen there depending on test order
* generate: back to plain symbolic binding, the static-1 rule wasn't worth it
* custom kernels: Q4_K support (ggml type 12)
Q4_K is Q5_K without the high-bit array: same d/dmin/scales layout (so
_q5_scales works unchanged), 144-byte blocks, qs at word 4. both the dp4a
decode kernel and the WMMA prefill kernel take a ggml_type branch now.
Qwen3-8B Q4_K_M: decode 16.5 -> 114.8 tok/s, prefill 69 -> 536 tok/s
* raise line count to 26500 (qwen did it)
* benchmark qwen3.8
* little updates
Stream 240KB chunks into two alternating 256KB SRAM bounce windows; each chunk
ends in a unique 512B sentinel that a prebuilt SDMA ring polls before copying
the chunk to VRAM, followed by an in-order drain fence that the host waits on
before re-arming a window. No timing assumptions in either direction: the
sentinel is in-stream proof of data landing, the fence proves a full drain.
Adds a small pooled async bulk-OUT layer to USB3 so staging the next chunk
overlaps the wire, and a slot_start parameter to scsi_write for the second
window. 107 -> 276 MB/s copyin on tinyc8 (Kryo-3XX host).
Co-authored-by: tiny <tiny@local>
* compile kernels in parallel (slop)
* cleanups
* fixes
* hand cleanups
* no PARALLEL with VIZ
* ugh, we need to remove the n from names
* no hcq2 in that test
* main for abstractions3
* fix compile server to be fork safe
* fix num cpu threads in context
* just use a lock
* fix __main__ on spawn
* compileonce
* python3 is double
* xdist sizing
* parallel compile
* fetch supports extract
* fix issues
* revert compiler server to master, drop compileonce
* normal workflow
* PARALLEL=0 for QCOMCL compiletest
geohot
sirhcmandGitHub