Replaces llama.cpp Gemma4-12B-QAT (payload + binary retained as rollback).
SGLang in a self-contained python3.12 venv with an extracted CUDA 13.0
nvcc toolchain for the sgl-kernel/flashinfer JIT.
Shape: fp8 KV, pinned 606,208-token pool / 4 x 151,552 slots, bf16 mamba
states pinned at 32 slots (the engine autofills the ssm pool to a fixed
~5.1 GB regardless of dtype), eager prefill (no headroom for graph
capture), no spec decode (the checkpoint ships no mtp.* tensors despite
mtp_num_hidden_layers=1 — NEXTN drafts with an uninitialised 4.24 GiB
ghost at 0% acceptance).
Verified: 83.9 tok/s decode, 170k single prompt, 4 concurrent 54k
requests, vision (in-checkpoint tower), reasoning_content field.
Five 262144-token contexts passed every boot gate (KV 1,350,182 / 5.15x) but
left 0 bytes free at runtime: a real client request (995-token prompt,
max_tokens=261149, sampling params straight from models.yml) killed Worker_TP1
with torch.OutOfMemoryError 'Tried to allocate 50.00 MiB ... 0 bytes free'.
Boot is not proof of runtime survivability.
Four slots x 262144 fp8 (KV_BYTES 8,311,013,376/rank) returns ~1.93 GiB per
rank for activations/graph capture.
Unit: Restart=always instead of on-failure — a runtime worker fault makes
vLLM's API server exit clean, so ExecStart returned 0 and on-failure never
fired, leaving the endpoint down until a human noticed. Also
StartLimitIntervalSec=0 and OOMScoreAdjust=-500.
Frank's call after the night ladder: five contexts at the model's full native
262,144 window beat six shortened ones. Six full-length contexts do NOT fit at
any KV dtype (6 x 241664 fp8 asked 10.70 GiB against 9.18 GiB free — the
6-seq graph/activation footprint leaves too little for the pool); 5 x 262144
at fp8 PASSED on the first rung: KV_BYTES 10,388,766,720/rank, measured
'GPU KV cache size: 1,350,182 tokens, Maximum concurrency for 262,144 tokens
per request: 5.15x' (39,462 tokens of margin).
Also documented: the SEQS=3 bf16 max (241,664, KV 732,359 / 3.03x) measured
earlier the same night, and the root-cause read that the 7.2 GiB KV ceiling
law was the MTP4 workspace, not VRAM. Seam and harness (contextWindow 262144)
re-pinned.
Night ladder result (Frank's subtract-16k method): 262144 and 245760 both OOM
at CUDA graph capture ('Tried to allocate 76.00 MiB ... 0 bytes free'), 237568
passed (KV 718,606 tokens), and the +4096 bonus step PASSED -> 241664 is max
possible and DEPLOYED at util 0.98 (KV 732,359 tokens / 3.03x concurrency at
241,664-token requests). Util stays 0.98: lowering it shrinks the engine's own
budget instead of buying slack.
Client contract restored (the GLM rollback d91379d dropped all three and the
qwen rollback inherited the gap): --reasoning-parser qwen3 fixes the malformed
thinking block (thinking was landing inline in content with a dangling
</think> and a null reasoning field), --enable-auto-tool-choice with
--tool-call-parser qwen3_xml makes tools work, and
--default-chat-template-kwargs enables enable_thinking + preserve_thinking.
Harness: contextWindow 241664 (compaction gap recomputed), preserve_thinking
moved into chat_template_kwargs.
The 0.98 boot died at KV allocation: torch OOM tried 10.28 GiB (the exact
3x262144 bf16 pool = 11,047,010,304 bytes) with 9.46 GiB free per rank.
The rootless bench probe had booted the same numbers; the root unit's
pinned-expert path holds ~0.8 GiB/rank more. Util 0.90 returns ~2.5 GiB/rank
of workspace with the pool ask unchanged; measured KV capacity law
(798,280 tokens at 11.05e9) unchanged.
Overnight redo campaign verdict + Frank's call: the fp8+MTP4 shape is retired
to git tag frank-primary-fp8-mtp-20261001. The new incumbent: KV cache
unquantized (auto), no speculative decode, KV_BYTES 11,047,010,304/rank
(engine-measured 798,280 tokens, 3.05x concurrency at 262,144-token
requests), MNBT 1024, MoE gather 32x64 (+18.7% decode, grows with depth).
Measured at the deployed shape: 39.38/38.79/36.42 tok/s (166/1087/4064 ctx).
Seam re-pinned: the 7.2 GiB/rank KV ceiling was the MTP4 workspace, not VRAM;
spec decode absent by measurement; bf16 trade documented (clean cache +
262k windows for ~14% decode at 4064 ctx).
- AGENTS.md: fleet table + frank-primary section back to qwen FP8 + MTP4
(37.4-40.6 tok/s, 54% acceptance); GLM runbook moved to the retired index.
- runbook: RETIRED verdict — DFlash2 drafter served and drafted but
acceptance pinned 1.00-1.38 on every text class including verbatim
repetition (mis-wired aux capture, not a weak drafter); 5x slower than
baseline at ~1.2 acceptance.
- log.md: 2026-10-01 rollback entry.
- deploy-remote.sh: LLM_PRIMARY_CONTAINER back to frank-primary-qwen;
glm-probe joins the legacy sweep list.
- frank_primary_unit.bats: re-pinned to the qwen shape — KV qualification
law (7.2 GiB/rank, 3.65x concurrency; 9/16 GiB OOM), fp8 q8/q8 KV
PRESENT, MTP4 PRESENT (3.1-3.4x over the 12.11 no-spec baseline), no tool
dialect, automatic offload. The old 'MTP buys zero' law carried the
prefetch tax and is retired.
DFlash2 experiment declared failed 2026-10-01: the drafter serves and
drafts (12.8 tok/s proposed) but acceptance is pinned at 1.0-1.38 on ALL
text classes including verbatim repetition — the aux-hidden-state capture
semantics are wrong (single-stream entering-state vs post-layer/comb is
the untested alternative). At ~1.2 acceptance the 8-token verify tax makes
decode 5x SLOWER than the 9.9-10.0 baseline (1.6-1.8 tok/s measured).
Restores the verified 37-40 tok/s shape: qwen3.8-flash-next FP8, MTP4,
prefetch off, 200k x 3 seqs, fp8 KV.
use_aux_hidden_state_outputs lives on the gpu_model_runner instance
(gpu_model_runner.py:623/678), not on the spec config, so the first
registry check never fired and the reserve stayed at 13,600 B/token.
Switch to method in {dflash, dspark, extract_hidden_states}.
The dflash boot cleared the unifier and the kpool tail (02/03) and died on
KV capacity: on the generic path the mamba states are NOT aliased into the
MLA pages (the glm5 tensor layout's whole point), so the capacity estimator
demanded ~30.6 GiB fixed against 0.92 GiB available (the plan's KV reserve
registry only knows the target's 13,600 B/token).
Patch 04 keeps the aliased glm5 pool and appends the drafter's uniform
group and per-layer hidden-state groups beside it: grouping partitions
hidden + draft specs, the layout returns a 10-tuple, the builder sizes one
shared backing allocation (pool + extras), and the memory estimator adds
the extras. Patch 04b extends the GLM registry kv_bytes_per_token by the
drafter (4,096 B/token) and hidden-state (5 x hidden x 2 B/token) costs so
the expert plan reserves enough.
Local smoke tests (CPU torch + real spec classes) validate grouping shape,
layout recognition, builder block math and memory accounting against
synthetic GLM+draft+hidden spec sets. Apply-verified: reproduces shas
5521e50b... / 1a623d5c...
Patch 02 (MLA-multiple max page) cleared the KV page-size unifier: the
dflash boot now reaches cudagraph capture. New blocker: the kpool decode
kernel asserts tail_kv_cache.shape[2] == pool_size; KpoolTailSpec pins
block_size == index_kpool, so the unifier's divisible branch must never
grow it. Patch 03 = 02 + a KpoolTailSpec branch that pads page_size_padded
instead. Apply-verified: reproduces patched sha 5093e4b7... Runbook
updated with the full KV grouping journey.
The v4 boot named the actual miss: model_runner.py:425 calls
set_eagle3_aux_hidden_state_layers(self.model) on the instance model_runner
LOADED -- and for GLM-5.3-Flash that is Glm5NextForConditionalGeneration (the
multimodal wrapper; APIServer log: 'Resolved architecture:
Glm5NextForConditionalGeneration'), which wraps the text model as
self.language_model = Glm5NextForCausalLM (glm5next/common/model.py:1113).
v4 put the three presence-checked attributes only on Glm5NextForCausalLM,
so the isinstance check on the conditional still failed with the same
RuntimeError.
v5 adds to BOTH wrappers, MRO-safely (attribute route, no Protocol base):
- Glm5NextForConditionalGeneration: supports_eagle3 / has_own_lm_head /
has_own_embed_tokens + the two methods delegating to self.language_model
(the pixtral.py:498 pattern);
- Glm5NextForCausalLM: the same three attributes + the two methods that
unwrap to self.model (the EagleModelMixin trunk -- v1's correct bodies).
Also proven by MRO simulation: the ForCausalLM's five-base ordering cannot
take SupportsEagle3 as a base in ANY position, so inheritance was never a
viable route for this model.
base 008eeaa2... -> patched bc9913a1..., apply-verified against the pinned
image (patch -p1 reproduces the sha256 and matches the mounted file).
The v4 window boot named the real presence-check rule: isinstance against a
runtime_checkable Protocol walks __protocol_attrs__ (which includes members
inherited from BASE protocols) and getattr_static()s each one off the instance,
breaking on the first missing attribute. SupportsEagleBase declares
has_own_lm_head/has_own_embed_tokens as plain instance-var annotations, so
v3's bare supports_eagle3 flag still failed. Replicated in-image with stock
typing: flag-only class -> False; flag + the two base members -> True.
v4 declares all three on Glm5NextForCausalLM:
supports_eagle3: ClassVar[Literal[True]] = True
has_own_lm_head: bool = False
has_own_embed_tokens: bool = False
(False is correct for the two: the wrapper delegates embed/lm_head to its
trunk. The MRO-safe attribute path is kept from v3.)
base 008eeaa2... -> patched 60f27aad..., apply-verified against the pinned
image (patch -p1 reproduces the sha256 and matches the mounted file).
v2 (inheriting SupportsEagle3 as a base) taught the real lesson: the isinstance
check PASSES and the class then dies at creation with
'Cannot create a consistent method resolution order (MRO) for bases ...'
-- GLM's outer wrapper already inherits vLLM Protocols (SupportsPP,
MixtureOfExperts, HasInnerState, IsHybrid) whose ordering is incompatible with
appending another Protocol base.
v3 avoids the MRO entirely: declare
supports_eagle3: ClassVar[Literal[True]] = True
as a class attribute on Glm5NextForCausalLM. Runtime-checkable Protocols do
presence-based isinstance, so the bare flag satisfies supports_eagle3(model)
without touching the MRO. The aux-layer plumbing stays on the trunk via
EagleModelMixin.
base 008eeaa2... -> patched 4c9a632b..., apply-verified against the pinned
image (patch -p1 reproduces the sha256 and matches the mounted dflash-patch
file byte for byte).
v1 hand-wrote set_aux_hidden_state_layers/get_eagle3_default_aux_hidden_state_layers
and the boot reproduced the identical 'Model does not support EAGLE3 interface'.
Reason: supports_eagle3() is an isinstance check against a runtime-checkable
Protocol whose members include a required ClassVar flag
(interfaces.py:1663 supports_eagle3: ClassVar[Literal[True]] = True), so methods
alone do not satisfy it. Inheriting SupportsEagle3 -- the idiom every model in
this image uses, including the fork's own DFlashLagunaForCausalLM -- supplies the
flag, SupportsEagleBase's has_own_lm_head/has_own_embed_tokens, and working
default implementations of both methods; the hand-written methods are deleted.
base 008eeaa2... -> patched 5eea89c5..., apply-verified against the pinned image.
Runbook: corrected method name, recorded the Protocol trap as the generalizable
lesson for future spec-decode ports.
Clarifies scope in the artifact header: torch.compiler.disable is itself a graph break,
so 05 helps only a hypothetical fullgraph=True route (where the workspace acquisition
would be better restructured out of the forward than made opaque). It is kept because
the discovery is real and the wrapper is a no-op when nothing is compiling, but it is
not required for the break-tolerant route whose only confirmed blocker is the dynamo
name-map KeyError.
Also verified in this pass: glm53-dflash-01-aux-hidden-states.patch applies cleanly to
the pinned image and reproduces its recorded patched sha256 (4ff6ff8d...), so all six
artifacts in scripts/patches are now apply-verified.
glm53-dflash-01-aux-hidden-states.patch: implements the interface whose absence makes
the engine refuse the drafter with 'Model does not support EAGLE3 interface'. Adds
EagleModelMixin to Glm5NextModel, captures the single-stream hidden state entering each
requested layer in the loop, returns (hidden_states, aux_hidden_states) as a tuple when
the list is non-empty, and exposes set_aux_hidden_state_layers /
get_eagle3_default_aux_hidden_state_layers on Glm5NextForCausalLM -- the contract read
out of interfaces.py:1694/1712, eagle3_utils.py:26-31, deepseek_v2.py:1551-1585 and
cudagraph_utils.py:621, with kimi_k25/pixtral as the delegation reference.
Independent of the compile changelist (01-05); it patches the same file as 01, so 01
goes first if both are applied. Scope is honest in the header: this makes the interface
present (the prerequisite for driving DFlash2), it does not promise draft quality and is
not a decode-speed fix.
Five runs of the changelist now reproduce the same single failure with breaks genuinely
tolerated (patch 01+02+03+04, VLLM_COMPILE_FULLGRAPH=0, VLLM_USE_AOT_COMPILE=0,
R4D_GLM_FUSE_QKV_RMSNORM=0): the trunk TRACES ('Dynamo bytecode transform time: 1.08 s')
and the boot dies on KeyError: 'in_proj_qkvbfg_a'.
What it is NOT, each ruled out by measurement rather than argument:
- not one of dynamo's four nn_modules lookups: /home/dank/dynamo-probe wraps three of
them and stayed silent over two runs, with the probe's install line confirmed in both
the API-server and worker processes;
- not a dynamo-internal failure: TORCHDYNAMO_VERBOSE=1 adds no internal stack trace;
- not a literal string lookup: the name appears tree-wide only as the KDA layer's
attribute (kda.py:217, used at :407) and in the load_weights shard table
(model.py:780-785); the four sites in the
quantization utils do compute exactly this bare key but every one is guarded
(), so none can raise it;
- not the pluggable-layer registry, which is keyed by class name.
What it IS: a plain Python KeyError raised inside the traced
Glm5NextDecoderLayer.forward, in a callee dynamo INLINED (frame chain
profile_run -> _dummy_run -> execute_model -> glm4_1v.py:2358 -> the compile wrapper ->
eval_frame.compile_wrapper -> Glm5NextModel.forward:748 -> module._call_impl ->
Glm5NextDecoderLayer.forward:456; the reported line is that method's def line, which is
how inlined frames surface).
Patch 03's header and the runbook now record the remaining instrument (TORCH_LOGS=+dynamo
and grep, or wrap the inlining candidates — the mHC CustomOps and the indexer op — with a
KeyError probe), and the two breaks that are real but non-fatal in this configuration
(patch 04's fused-norm launcher cell, patch 05's workspace acquisition).
tests/run.sh: 362 seams pass.
Root-caused the moving failures: they were my own scratch launcher. Its compile
branches were restructured and the '-e VLLM_COMPILE_FULLGRAPH=0' line was dropped, so
four consecutive boots mounted the wrapper patch while _FULLGRAPH still defaulted to
True — and with fullgraph=True EVERY break is an error, so each boot reported whichever
break came first (cell -> workspace -> compiler-disable). I had blamed
TORCH_LOGS=graph_breaks for the difference; that was wrong, the flag only logs. The
launcher sets the env var again, and patch 04's header, the runbook and the work log
now carry the true run history as a table instead of the bad theory.
New artifact: scripts/patches/glm53-compile-05-workspace-acquisition-opaque.patch —
the indexer acquires its shared workspace INSIDE the forward (its profile-run branch
literally says 'Reserve workspace for indexer during profiling run'), and
v1/worker/workspace.py:_ensure_workspace_size reads ContextVar/ubatch state, which
dynamo cannot trace. Wrap both call sites in a torch.compiler.disable'd helper
(narrower than decorating the workspace manager, which every model would go through;
a no-op when not compiling).
All five artifacts verified against the pinned image: applied in sequence with patch
-p1 they reproduce their recorded patched sha256 exactly (6e9cb81b…, 0751852e…,
732363bb…, 84969993…, c154115a…).
tests/run.sh: 362 seams pass.
Working the dyno name-map blocker one step at a time, with the tools to see it:
- All four patch artifacts are now genuine diff -u patches (they were prose + pseudo
hunks before, useless to patch(1)). Verified against the pinned image: each applies
clean and reproduces its recorded patched sha256 (6e9cb81b… / 0751852e… / 732363bb…
/ 84969993…).
- New: scripts/patches/glm53-compile-04-fused-qkv-rmsnorm-optout.patch — takes GLM's
MLA block off the fused-norm JIT kernel (R4D_GLM_FUSE_QKV_RMSNORM=0; the flag is
already a constructor arg glm5next sets at attention.py:583). Reason: in
models/common/ops/fused_qk_rmsnorm.py the kernel object is bound at line 223, AFTER
its use at line 211 — the launcher family (@kernel_launcher, 39 call sites) is what
breaks dynamo's break analysis.
- New diagnostics, both reusable and recorded: TORCH_LOGS=graph_breaks +
VLLM_USE_AOT_COMPILE=0 to name each break with its user stack, and
/home/dank/dynamo-probe (sitecustomize.py on PYTHONPATH, VLLM_DYNAMO_DUMP_ATTRS=1)
to print the dynamo name-map miss detail. Gotchas: a .pth on PYTHONPATH is NOT
executed, and the scratch launcher reset ENVARGS (now ${ENVARGS:-}), which silently
dropped env vars and wasted a boot.
Measurement discipline correction: I first recorded the 'Read uninitialized cell' and
'Unsupported method call' failures as fatal blockers. They are not established as
such — they appear only when TORCH_LOGS=graph_breaks is set, while a flag-free
break-tolerance run gets past break handling and dies on the name-map KeyError. The
likely mechanism is that graph_breaks logging makes dynamo build a break
*explanation*, and that analysis itself trips on the uninitialized cell. Patch 04's
header, the runbook and the work log now carry the confound explicitly rather than
the overclaim.
tests/run.sh: 362 seams pass.
DFlash2: fetched incoai/GLM-5.3-Flash-DFlash2 (2,342,169,800 B, sha256 matches the
Hub LFS oid) and tried --speculative-config {"method":"dflash","model":"/draft"}.
Config and drafter are accepted; the worker dies with 'Model does not support
EAGLE3 interface' - so the only gap is the target-side aux-hidden-state contract
(set_aux_hidden_state_layers + per-layer capture + the two-value forward return,
as deepseek_v2.py demonstrates twice in this image). Recorded the exact interface,
the GLM-specific unknown (what the single-stream state entering layer k means for
an mHC multi-stream trunk) and the honest expectation: no speedup on a
token-proportional pointwise bound.
fp8 KV: recorded the deciding measurement - the fork's KV reserve model is
dtype-blind (the fp8 boot printed the identical 1760 MiB / 13600 B-per-token plan),
so fp8 cannot buy capacity even if the gfx12 MLA path were rewritten. Verdict: do
not spend effort there.
Compile blocker narrowed from 'a latent GLM bug' to dynamo's name map: the
KeyError is thrown by torch/_dynamo/output_graph.py's nn_modules lookup (:3000,
registered :1259/:1384, _modules[..] prefixes stripped :1225), i.e. the compiler
cannot resolve a submodule by name while tracing the decoder layer. Next step is
TORCHDYNAMO_VERBOSE=1, with the note that the scratch launcher resets ENVARGS.
tests/run.sh: 362 seams pass.
A torch-profiler kernel-time window (496 decode tokens, rank0 trace parsed for
device-activity time) on the deployed GLM-5.3-Flash shape falsifies this page's
own "the expert gather is the bound" explanation:
46.18 s of kernel time / 496 tokens = ~93 ms per token (actual step ~104 ms)
34.8 % vectorized_elementwise_kernel 708,500 launches (1,428/token)
10.0 % r4d_gemm_moe_w4a16_nt_b16_m64 (the MoE GEMM)
6.7 % elementwise_kernel_manual_unroll
6.4 % wvSplitK_hf_sml_ (split-K GEMM)
8.9 % router: sbtopk + warpMergeSortTopK + bitonicSortKVInPlace
6.5 % r4d_mhc_* (hyper-connections)
1.05 % r4d_moe_lru_gather_kernel <-- the retracted suspect
The earlier arithmetic was circular (it assumed the gather consumed the whole
step). The real bound is unfused pointwise glue: the fork compiles its own Qwen
trunk (support_torch_compile in models/qwen4_exp/amd/model.py) and not its GLM
trunk, so the engine boots with "provides neither a compiled submodule nor
breakable CUDA graphs. Overriding to FULL_DECODE_ONLY." That also explains why
MTP4 measured flat — pointwise cost scales with the token count, so spec decode
amortizes nothing here — and it retires arena/hit-rate tuning (the gather is 1 %).
Staged as documented patch artifacts (not wired into any serve script; the
engine does not yet boot with them) — scripts/patches/glm53-compile-0{1,2,3}*.patch:
-01 models/glm5next/common/model.py decorate Glm5NextModel (R4D_GLM_COMPILE)
-02 model_executor/kernels/r4d_lib.py freeze 10 capability probes that do
os.path.exists/.so lookups inside the traced forward. functools.lru_cache
does NOT protect them: dynamo re-traces the wrapped body even when warm
(verified in-image); a plain module-global read compiles.
-03 compilation/wrapper.py make fullgraph=True opt-out, because
suppress_errors does NOT cover graph breaks and this trunk can never
satisfy fullgraph (indexer metadata syncs to host — .item() at
glm5next/amd/sparse_indexer.py:263,419).
With all three the trunk traces (Dynamo bytecode transform time 1.08 s) and dies
on KeyError: in_proj_qkvbfg_a in the mHC/KDA path — a latent bug the compile
exposes, and the remaining blocker.
Also records two structural constraints: no new HIP kernels are possible (r4d.so
ships without sources), which binary-blocks the fused router at GLM's 288
experts; and DFlash2 for GLM is a TARGET-side port, not a drafter port
(incoai/GLM-5.3-Flash-DFlash2 is architectures ["DFlash2DraftModel"] +
model_type qwen3, already mapped at registry.py:636; glm5next lacks only the
aux-hidden-state plumbing deepseek_v2.py demonstrates).
Runbook frontmatter, intro, mermaid and wall section, AGENTS.md and the work log
corrected to match the measurement. tests/run.sh: 362 seams pass.
The deployed unit failed its first boot with the engine core OOMing during
profile/graph capture: "Tried to allocate 662.00 MiB. GPU 0 has a total
capacity of 31.86 GiB of which 240.00 MiB is free ... try setting
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True". The expert-offload plan
fills VRAM to its modelled headroom by design, so the default allocator's
fragmentation leaves the capture phase nothing — every successful bring-up
probe (both scratch-port and 7770) carried this env, and the repo recipe
dropped it.
Also pins the seam: the setting is part of the qualified shape, not an
optional tuning.
Payload: Intel/GLM-5.3-Flash-W4A16-AutoRound (AutoRound symmetric int4 group
128, auto_round:auto_gptq, 34 shards + 2 extras, 113,074 tensors, ~170 GiB).
It is the only W4A16 checkpoint this fork can serve with expert offload: its
quant_method "auto-round" is overridden to "inc", which routes the MoE to
libr4d's grouped W4A16 GEMM (R4dW4a16MoEMethod). A compressed-tensors W4A16
build of the same model resolves to TritonWNA16Experts, which has no offload
recipe, and refuses to serve at all (170 GiB of experts cannot fit 64 GiB of
VRAM, so offload is mandatory).
Shape: tcclaviger RDNA4 vLLM fork (same digest-pinned image as the 2026-09-29
qwen cutover), TP=2, 131,072 ctx x 2 seqs, bf16 KV (fp8 is ILLEGAL here — it
disables the sparse MLA backend's rope-free Triton path, the only MLA decode
gfx12 has), util 0.98, NBATCH 2048, --enable-expert-offload
--expert-precise-memory, prefix caching, glm47 tool + reasoning parsers,
MTP and expert prefetch absent by measurement.
Measured: decode 9.9-10.0 tok/s, prefill 918 tok/s (warm), arena 14.51 GiB =
2,383 slots/rank over 12,096 host-backed experts (73.64 GiB pinned/rank).
THE 35 tok/s TARGET IS NOT MET and the shortfall is measured: live telemetry
(R4D_LRU_TELEMETRY=1) gives a 0.53-0.66 per-layer hit rate, a step moves
0.82 GiB of misses in ~100 ms = 8.2 GB/s effective against a link measured at
22.74 GB/s — the same structural gather-path gap the qwen campaign found.
Growing the arena 13.08 -> 14.51 GiB changed decode by nothing.
Rejected by measurement: MTP4 (head loads, acceptance 3.76-4.90, but decode
unchanged and prefill collapses 918 -> 255 — a verification forward gathers the
union of the batch's experts), NBATCH 8192 (identical prefill).
Also: scripts/patches/gpu_worker-kv-cap.patch (fork-local fix for a KV-reserve
model that under-counts GLM's per-sequence KDA state; not needed by this shape),
runbook, AGENTS.md row + section, both indexes, work log.
Rollback: 2339fae (qwen FP8 + MTP4).
The classic 'podman images | grep -q' pattern races under set -o pipefail:
grep exits on its first match, podman dies SIGPIPE (141) writing the
remaining RepoDigests lines, and pipefail promotes 141 to the pipeline
status — the gate reads 'no match' and refuses to deploy (2026-09-30,
dankserver). Same bug in the serve script (it would restart-loop the
same way). Fix: capture podman's output into a variable first, grep the
variable. Patched in the R9V September deploys by luck of timing.
Unquoted value + set -a sourcing parses 'qwen-fp8' as a command ->
ExecStart exit 127 -> unit restart loop (caught on the first deploy).
The deployer's wait_for_health was killed and will re-run on the fresh
deploy.
Replace the R9V radiance fork with the tcclaviger RDNA4 vLLM stack (the
digest-pinned image the 2026-09-29 tuning campaign measured), serving
orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 (131 shards, 172.8 GiB,
already on disk) with MTP4 spec decode.
Qualified by boot probe on 2026-09-29:
- MAXLEN=200000 x SEQS=3 (600k total) at --kv-cache-dtype fp8 (q8/q8 KV):
'GPU KV cache size: 729,545 tokens; Maximum concurrency for 200,000
tokens per request: 3.65x' at KV_BYTES 7.2 GiB PER RANK.
- KV_BYTES is per-rank: global 16G and 9G both OOM (worker tried to
allocate the full size; ~7.5 GiB/rank is the ceiling with the arena).
- prefetch OFF by law (telemetry: 36% precision, ~50% wasted traffic);
TunableOp off (corrupts numerics).
- decode at this shape: 42.94/34.55/36.09 tok/s; arena 6,032 slots.
Trio: frank-primary.env (all FRANK_PRIMARY_* keys), serve-frank-primary.sh
(direct podman run; entrypoint execs vllm serve; waits on the container),
frank-primary.service (root unit, same ExecStop law, new container name
frank-primary-qwen). Deployer gates rewritten (image digest + shard count
+ weight-map index). download-frank-primary.sh is now inventory/verify
(repo is GATED — 401 anonymous, masked LFS hashes; payload already on
disk). Seam rewritten for the new contract. Pre-cutover tag:
frank-primary-r9v-q8-20260929.
The disk was at 93% with 137G free. Nine retired payloads are now moved
to ~/.AGENTTRASH/20260928-retired-models/ with a MANIFEST.md giving a
restore command per item — moved, not deleted, and nothing is freed until
the purge is approved. Every one maps to a runtime already retired in the
runbook index.
Only the .gguf was taken from Qwen3.8-27B-UD-Q8_K_XL/: the live
frank-large DFlash2 drafter lives in that directory and stays put. The
manifest warns against restoring the directory wholesale.
Kept: checkpoints/ (23G) — unreferenced by the repo, but ComfyUI is live
and its workflows were not searched, so it is not confirmed dead. Also
logs/, because x4e-run.log is a SPARSE file (3.3G apparent, 4K real) and
frees nothing.
The reference check caught a live bug: serve-frank-small.sh's fallback
MODEL_DIR still pointed at the retired GSQ-RCO payload frank-small served
on 2026-09-16/17. frank-small.env has always overridden it, so the drift
was invisible — but a missing or unreadable env file would have turned a
config problem into a "missing asset" naming a model that is no longer
there. Repointed at Gemma4-12B-QAT/, which the env file already sets.
Not taken, by request: podman images (279G reclaimable, 95%) and the
uv+pip caches (73G). Together 352G of regenerable garbage, larger than
what was reclaimed here.
The 29.3G payload rollback target went with the rest, so the payload-only
rollback is now a re-fetch first; the runbook and AGENTS.md say so.
The ad-hoc status commands kept producing wrong readings, most seriously
reporting "MemAvailable 0.0 GiB" on a host that actually had 240 GiB free.
Cause: /proc/meminfo separates key from value with a TAB, so
grep MemAvailable /proc/meminfo | tr -s " " | cut -d" " -f2
returns an empty field and awk prints 0.0. A false "host is out of memory"
reading is worse than no reading, because it invites an unnecessary abort.
dankstatus.sh parses on a colon split (tab-agnostic), cross-checks
MemAvailable against MemTotal before printing either, and refuses to print
a zero value. It also pins the rocm-smi CSV column order (device, total,
used) which was previously read swapped, and takes only the first line of
systemctl is-active/is-enabled, which emit more than one line here.
Incidental finding while validating: rocm-smi reports card2 (the 7700X iGPU)
as 0.50 GiB total, not the 123.0 GiB torch.cuda reports. torch was exposing
the GTT aperture. The device is smaller than documented, not larger.
Tests: 346/346 pass (no test surface touched).
Public, ungated repo so no auth header is sent — a token on the curl
command line is world-readable via /proc/<pid>/cmdline. Size and sha256
are the Hub LFS values (31457990272 / 9076f99a…), verified after fetch.
Not wired into frank-large.env: the gufo load test has to pass first.
gufo is a vertical inference engine built for exactly this hardware
(Ryzen AI MAX+ 395 / Radeon 8060S, gfx1151), and Dankclaw's frank-large
has swapped runtimes seven times in search of a better one. This one
ships no prebuilt Linux binary, so the runtime is BUILT FROM SOURCE on
the host and the payload, port, alias and client contract stay exactly
where they are.
Runtime: gufo v0.1.1 (b0f8673) at /home/dank/GitPackages/gufo-v0.1.1,
built by the new scripts/build-gufo.sh. It links the host ROCm 7.2.4
through its RPATH — the opposite of the unsloth prebuilt it replaces,
which carried its own — so two host workarounds are encoded in that
script: Ubuntu's librocwmma-dev installs a truncated rocWMMA header set
that shadows AMD's and cannot be overridden with -I in HIP mode, and the
ROCm clang prefers Ubuntu 26.04's GCC 16 headers over the GCC 15 whose
libstdc++ is actually installed.
Shape is unchanged: one session, full 262144 ctx, the same UD-Q8_K_XL
target and the same DFlash2 drafter, greedy, text-only. Speculation
moves from llama.cpp's fixed depth 4 to gufo's adaptive block policy
capped at 7 proposals.
Client-contract deltas, all documented: the chat template is compiled
into the binary (no --chat-template-file, so the Qwen-Sharp 73k-token
failure mode is gone), reasoning_effort now takes the full OpenAI set
(so models.yml's reasoningEffortMap fold is no longer required — kept
anyway, so this commit does not also move every high role from xhigh to
high effort), and there is no KV-quant or prompt-cache-RAM flag.
The test seam now runs the serve script against a stub engine and
asserts on the argv it actually execs, instead of grepping the script's
text. The deploy handler gates on the runtime binary and returns before
it stops the serving unit.
Measured trade, from gufo's own Q8_K_XL tables against llama.cpp
b11069: prefill +37% to +80%, cold load 4.5x faster, decode +7% at zero
depth — but -14% at 65,536 tokens of depth and -24% at 131,072. On the
8-bit target the crossover is ~32k, and frank-large runs one 262,144
conversation. The Q4_K_XL target shows no such regression.
download-frank-primary-fp8.sh + the fp8 manifest fetched the vLLM 0.30
payload that never deployed; the R9V cutover's payload is the R9V setup
flow's (packages + PLE, state-gated). Scripts move to old/ per policy.
Replace the vLLM 0.30 UVA-offload wiring (never deployed; llama.cpp
b11030 endpoint replaced by Frank's order, downtime window open) with
the R9V radiance-vLLM shape on the same root unit:
- serve-frank-primary.sh: layer the on-host qualified setup state
(104 tuned R9V_* keys, /home/dank/r9v-state/uncensored/setup.json)
under repo production overrides, then exec the fork's own launcher
through the docker->podman shim and block on docker wait so a crashed
engine is a nonzero ExecStart (launch.sh --detach alone would fake
green). Gates: pinned image by OCI manifest digest in RepoDigests,
PLE at pinned byte count (stat -c %s), all GGUF shards + mmproj.
- frank-primary.env: thin production overrides only (LAN bind 0.0.0.0,
:7770, alias frank-primary, container frank-primary-r9v, GPU BDFs);
tuned engine shape stays in the setup state by design.
- frank-primary.service: same root-unit shape (User=dank,
LimitMEMLOCK=infinity for the tiered-expert pin), plus ExecStop
podman stop — docker wait cannot stop the container child.
- deploy-remote.sh deploy_llm_primary: state-driven gates (state ready,
image digest present in rootless storage, PLE bytes, shard files)
replace the fp030 shard-index gate; false-green MainPID check kept.
- tests/frank_primary_unit.bats: seam rewritten for the R9V shape.
Unit restart SIGKILLs conmon before --rm cleanup runs, leaving the container
name registered: every restart then dies on name-in-use while the orphan
holds the port and GPUs (observed live 2026-09-27, NRestarts=6 with the
container Up and the unit auto-restarting). The serve script now clears its
own name before exec podman run; the deployer stops the unit and removes
the container BEFORE waiting on the port.
HOME inside the container pointed at the root-owned WORKDIR /app, so
vLLM's registry cache write hit PermissionError; HOME now points at the
cache volume.
The vLLM cutover itself is blocked on a structural wall (AMD Qwen4Exp impl
creates the 51.2B PLE table as a bf16 on-GPU param: 47.69 GiB transient on a
31.86 GiB card; the pinned-host UVA path that fits this hardware is
CUDA-gated). llama.cpp is restored live at :7770; host HEAD unchanged.
The llama.cpp b11030 trio becomes the vLLM 0.30 trio ahead of the
FP8 cutover on :7770 — same port, same alias, no client moves.
Serve shape: locally built ROCm gfx1201 image
(localhost/frank-primary:vllm030) execs via podman as dank from the
root unit (LimitMEMLOCK=infinity legalizes the pin). UVA offload:
--offload-backend uva pins the PLE family (47.75 GiB, 138 tensors:
131 n-gram table shards + 7 per-layer conv/proj) + routed experts
(114.86 GiB, 150,528 tensors) in host RAM — 162.60 GiB combined,
~81.3 GiB per rank under TP=2, hence --cpu-offload-gb 82 (80/card
strands 2.6 GiB of experts on the cards). Parsers per the host's
own Ornith-era precedent: qwen3_xml tools + qwen3 reasoning (vLLM
0.30 emits the channel as `reasoning` — reasoning_content is
renamed server-side) + enable_thinking/preserve_thinking template
kwargs. Spec decode stays env-wired and OFF at bring-up.
Deployer deploy_llm_primary rewritten: image gate, index-driven
shard-count gate (131), stale-container cleanup, port wait, and the
false-green hardening — health is only green when the unit's own
MainPID is live behind the answer (2026-09-26 lesson).
Seam tests rewritten to the vLLM contract (15 gates): offload
arithmetic, parser wiring, no-llama.cpp-ism, deployer gates. Suite
349/349 green.
scripts/manifests/qwen3.8-flash-next-fp8.manifest: 144 entries (131
safetensors shards + aux) with Hub LFS sha256s for
Qwen/Qwen3.8-Flash-Next-FP8. download-frank-primary-fp8.sh wraps
hf-fetch.sh: idempotent fetch, full sha256 re-verify of an on-disk
payload (verified live on dankserver: all 144 files match, ~10 min).
Part of the frank-primary vLLM cutover prep; the serve trio itself
lands with the image deploy.
The stock SearXNG image ships a native kagi engine; enabling it needs an
API key that must never enter git, and SearXNG does not expand env vars
in settings.yml (the live secret_key literally was '${SEARXNG_SECRET}'
until now — the app tolerated the placeholder string silently).
- settings.yml: kagi engine enabled (weight 2, kg shortcut, paging off so
exactly one API credit per aggregated query, timeout 8.0 vs 5.0 scrapers
so a slow paid query is not a wasted credit); api_key stays ${KAGI_API_KEY}.
- new scripts/searxng-render-settings.py: deploy-time renderer, allowlisted
to KAGI_API_KEY/SEARXNG_SECRET/SEARXNG_BASE_URL, fails the deploy loudly
on an unresolvable placeholder, idempotent on redeploys.
- deploy-remote.sh copy_searxng_config runs the renderer against
/opt/searxng/.env after copying settings.yml into /opt/searxng/config.
- KAGI_API_KEY placeholder added to dankclaw/.env.example (real key staged
into the host-side /opt/searxng/.env, mode 600, out of band).
- bats: engine wiring, renderer substitution, loud-failure, deployer seam.
- docs: SearXNG.md Kagi section + renderer law; Providers.md carve-out on
the no-credentials policy (paid engine acceptable when bounded + ordered).
Verified live on dankclaw: direct API probe via the gluetun netns (200,
data.search[] shape), then an aggregated query returned 29/29 results
tagged kagi, zero unresponsive engines.
The first no-redirect HEAD fix read X-Linked-Size on every 3xx, but
non-LFS files on a Xet-backed repo answer 307 to HF's own
/api/resolve-cache WITHOUT any X-Linked-Size header. The fallback to the
stub's Content-Length recorded the redirect page size (254-264 bytes)
instead of the file size, so hf-fetch.sh failed the size check on every
aux file it verified against the regenerated manifest (README.md
"65400 != 254", config.json "72514 != 258", 7 files on the FP8 repo).
Distinguish the two redirect shapes: 3xx WITH X-Linked-Size = LFS/Xet,
whose X-Linked-* headers are the truth and whose redirect must NOT be
followed (the Xet bridge ETag is the CAS hash, not the sha256); 3xx
WITHOUT = non-LFS, where the Location (followed once, via urljoin)
answers 200 with the real Content-Length and a git-blob-sha ETag that
the 64-hex gate correctly rejects (size-only check for those files).
Verified against Qwen/Qwen3.8-Flash-Next-FP8: all 7 previously-failing
aux files now record their on-disk byte counts exactly (README.md 65400,
config.json 72514, merges.txt 3353259, ...), shard shas unchanged
(model-00002 6841fe21...), total 172.82 GiB across 144 files.
On Xet-backed repos (Qwen/Qwen3.8-Flash-Next-FP8), the resolve endpoint 302s
to a CDN bridge whose final ETag is the Xet CAS chunk-dag hash, NOT the
content sha256. urllib follows redirects, so the manifest recorded the Xet
hash and every downloaded shard failed sha256 verification against a hash no
download could ever match (8/8 shards FAIL, all sizes correct, on-disk bytes
matching the Tree API's lfs.oid). Classic LFS CDNs serve ETag=sha256, which
is why this was invisible on prior fetches.
Fix: HEAD via http.client, which does not follow redirects. 200 = direct
answer (Content-Length + ETag, non-LFS); 302 = LFS/Xet redirect
(X-Linked-Size + X-Linked-ETag = content sha256). Verified against the FP8
repo: shard model-00002-of-00131 now records 6841fe21..., the sha256 of the
bytes already on disk (cross-checked against the Tree API lfs.oid).
Instrument the live server to answer "can -ot target the most active
layers": AMD IBS op sampling on llama-server during 4 mixed concurrent
generations, sampled load addresses mapped through /proc/<pid>/maps to
backing tensors with extents derived from consecutive GGUF header
offsets (no quant type-size tables - a wrong per-element size silently
mis-buckets tensor tails, first run lost 59% of expert samples to a
junk bucket).
Measured (75 s window, 2.7M expert-tensor samples): per-block read
volume is FLAT, 0.93x-1.08x of mean across all 40 CPU-resident blocks.
MoE top-k routing fires every block every token, so "most active
layers" do not exist at -ot granularity - any 8 GPU-resident blocks
are equivalent, the lever is COUNT, already at the 5/3 VRAM ceiling.
Real skew is per-EXPERT (top 10% of 48x512 cells carry 40.5% of
samples), below tensor-override granularity. No config change.
scripts/frank-primary-expert-activity.py: dump + sample + analyze in
one entry point; --analyze-only re-analyzes saved samples. Instruments
production serving in place; not a benchmark.
Blocks 0-4 on ROCm0 + 45-47 on ROCm1. This exact shape was probed and
battery-verified earlier the same session: card0 2.54 GiB free at boot,
2.11 GiB after the vision graph (steady under load), card1 unchanged
at ~3.3 GiB. Card1 cannot take another block (0.82 GiB would OOM the
vision graph + boot drift), so 5/3 is the physical limit at the
4x262144 / V-q8_0 shape. OT assertion pins the new value.
- CONTEXT_SIZE=1048576 / PARALLEL_SLOTS=4: four conversations, each the
full native 262144 window (per-slot window unchanged; slots divide
the pool, they never multiply it).
- CACHE_TYPE_V=q8_0 / CACHE_TYPE_K=f16: main KV drops 24,576 -> 18,816
B/token (-23.4%); the shared MTP drafter's KV ignores the V type
(measured: 2,048 MiB draft pool = the f16 law at 1,048,576 cells).
- OT_OVERRIDES: blocks 0-3 on ROCm0 and 45-47 on ROCm1 stay
GPU-resident (2.42/2.49 GiB per block, measured); the rest stay
CPU-computed. Measured free under load: ~4.9 GB card0 / ~3.4 GB
card1; a 5th block on card0 lands at 2.27 GB, under the 3 GB
checkpoint floor, so blocks are atomic and 4 is the safe fill.
- serve script: --cache-type-k/v env-wired tunables; ONE joined
--override-tensor flag (b11030 deprecates repeated flags, and the
loader is first-match-wins, so the specific patterns must precede
the broad exps=CPU sweep).
- seams: the window test asserts per-slot = 262144 at 4 slots; new
KV-type test; OT exact-value assertion pins the ordering law.
Both b11030 serve scripts never passed --metrics, so /metrics answered
501 and Prometheus could not scrape frank-large (silently down since
its deploy) or frank-primary. frank-primary also gets -lv 4 (VERBOSITY
env tunable, default 4) for the full per-buffer load breakdown.
The shared MTP head borrows the main model output head, which a -ts 1,1
layer split places on ROCm1. With tensor overrides set, n_copies==1 in the
draft's scheduler, and a borrowed output.weight in a ROCm1 buffer aborts the
draft graph (ggml-backend.cpp:941) when the draft context is on ROCm0 —
the unit crash-looped 4x at ~12 s per attempt during the 2026-09-26 cutover.
Probed live on the host: ROCm0 aborts identically, ROCm1 loads and serves.
Discovery, not guesswork: dump the GGUF v3 headers of the fetched parts
(scripts/dump-gguf.py, added — a dependency-free reader that never
touches tensor data, so a 50 GiB part scans in milliseconds) before
the first load.
What the dump says (parts 2-3, 2026-09-26):
- The 51B-param PLE n-gram table is ONE tensor: per_layer_token_embd.
weight, dims [160, 320001536], Q8_0, 54.4 GiB — exactly the whole of
part 3. It is the same tensor the shared MTP head borrows, so the
=CPU override is required for placement AND for the draft load
(GPU-resident borrow: "invalid vector subscript").
- The model is qwen4exp, 512 experts, top-10 routing
(qwen4exp.ple.ngram_size=3, heads_per_ngram=8).
- Part 1 is metadata-only (ntensors=0); parts 2-6 carry tensors.
The placeholder pattern ngram_embeddings matched nothing — dropped from
OT_OVERRIDES (env, serve-script default, seam, AGENTS.md, runbook,
LLM index) before it could masquerade as a working override at load
time. Deployed set: exps=CPU,per_layer_token_embd.weight=CPU.
Suite: 346 green.
The 2026-09-26 same-day cutover that ended the Lucebox DS4.1 detour:
Dankserver's :7770 endpoint is now unsloth llama.cpp (the b11030-mix
prebuilt, ROCm gfx120X, bundled runtime) serving
unsloth/Qwen3.8-Flash-Next-GGUF at Q8_0 — 175.30 GiB across six sha256-
verified split parts — on both Radeon AI PRO R9700s, as a ROOT unit (the
frank-large shape: no container, no generated unit).
The deployed contract:
- Placement law: layer split -ts 1,1 with -ngl 999; the routed experts
and the 51B-param n-gram (PLE) table are COMPUTED in host RAM via
-ot exps=CPU,ngram_embeddings=CPU (the x4-pair measurement: streaming
over PCIe loses at 3.7 t/s, computing in RAM does not care).
per_layer_token_embd.weight=CPU is required by the shared MTP head's
borrow path (a GPU-resident borrow crashes the draft load).
mmap stays on; --cache-ram 16384.
- Spec chain draft-mtp,ngram-mod: the shared Q8_0 MTP head
(mtp-...-shared-Q8_0.gguf, 2,786,568,256 B) at depth 2, with the
draftless ngram-mod matcher chained behind it — the only legal way to
layer two spec implementations in one process.
- Vision explicit: --mmproj names the BF16 projector (this build never
autoloads; a server without it comes up healthy and text-only).
- Reasoning: --reasoning-format deepseek → reasoning_content; the
Qwen3.8 template's {xhigh, medium, low} reasoning_effort gate folded
by reasoningEffortMap in models.yml.
- One stream, the model's full 262,144 context; greedy server default
with Qwen-recommended sampling riding compat.extraBody.
The deployer's deploy_llm_primary is rewritten for the root unit: it
gates every artifact at its declared byte count (stat -c %s, never wc -c
— which read the whole 377 GiB lucebox file for minutes on 2026-09-26),
retires the lucebox user quadlet to dated home trash, waits out port
7770, relinks/enables/restarts the root unit and waits up to 1800 s for
/health. The quadlet-era tree under dankserver/frank-primary/ is
archived (dankserver/.AGENTTRASH/20260926-llamacpp-cutover/ on this
box; git history everywhere else) and its planner mapping replaced by
the scripts/ trio.
Observability flips back to the llamacpp namespace: a frank-primary
Prometheus job at 192.168.0.235:7770 (cross-host LAN — the netavark
published-port quirk that forces host.containers.internal for the two
Dankclaw endpoints is same-host only), the LlmEndpointDown runbook
text, and the dashboard description for three llama.cpp endpoints.
Docs: the new runbook (frank-primary-llamacpp-qwen38-flash-next-q8.md),
RETIRED banners on the lucebox and radiance runbooks and their index
entries, the AGENTS.md fleet section rewritten for the root-unit shape,
and the homepage frank-medium card renamed to frank-primary with the
Q8_0 Flash-Next spec (docs.html #models and the modality note updated:
images on frank-primary and frank-small, audio on frank-small only).
models.yml: the dankserver provider is "llamacpp (dankserver)" serving
frank-primary, contextWindow 262144, maxTokens 65536 (compaction-gap
law: 183,501 + 65,536 <= 262,144), input [text, image].
Suite: 346 green (frank_primary_unit.bats rewritten for the root unit —
16 tests pinning the placement law, the shared-head requirement, the
explicit projector, HIP env hygiene (BIND_HOST, never bare HOST), and
the deployer's byte-count gates; the quadlet-era planner and
changed-tests mappings rewritten for the trio; the observability scrape
test asserts three llamacpp jobs again).
Rollback: the lucebox image (localhost/lucebox-rocm:gfx1201) and the
378 GiB payload stay on the host; the Qwen payload stays fetched, so a
re-swap in either direction is a redeploy, not a refetch.
Frank
Comrade Chief Designer