Module gdn_flags 

GDN / SSM decode-path flags, resolved ONCE from the serve command line.

These three select KERNELS on the GDN decode path, and they are coupled: the FP16 h-state twins only exist on the fused-norm arm, so h_f16 without fused_norm reaches an FP32-only kernel that would read the FP16 pool as FP32 — plausible numbers, silent garbage. That coupling is checked at serve time by spark-server’s arg validation, not discovered at the first decode step.

Why these are set, not read

They were three independent std::env::var reads scattered across six call sites, each with its own convention (ATLAS_SSM_H_FP16 presence-gated — where =0 meant ON — and the other two == "1"). That is how the same flag came to be decoded two different ways in one binary. They are now ONE cell, written once from set_from_cli before any model is built.

The environment variables remain honoured when the setter never runs (a test, a microbenchmark example, an older script), so nothing that worked before stops working; the CLI wins when both are present.

Follow-up: this is process-scoped, so a hot-swap to a model with a different recipe keeps the first model’s kernel selection. The proper home is ModelLevers, which is carried per model — deferred because the h-state dtype is read from SsmLayerState construction sites that have no ForwardContext.

Structs

GdnFlags

Constants

MAX_F16_TWIN_DFLASH_GAMMA

The largest --dflash-gamma whose verify width still has an FP16 twin.

MAX_F16_TWIN_K

Widest chain-verify K with an FP16 h-state twin (gated_delta_rule_wy{5..16}_f16).

Functions

default_dflash_gamma

The served DFlash gamma for a drafter of this trained block size, when no --dflash-gamma was given.

flags

The resolved flags, falling back to the environment on first touch.

gdn_fused_norm_enabled

--gdn-fused-norm (legacy ATLAS_GDN_FUSED_NORM=1).

set_from_cli

Publish the command line’s resolution. Call once, before the model builds.

ssm_batched_recurrent_enabled

--ssm-batched-recurrent (legacy ATLAS_SSM_BATCHED_RECURRENT=1).

ssm_h_dtype_bits

SSOT decode of --ssm-h-dtype into the two h-state bits it publishes: (h_f16, h_f16_pool).

ssm_h_f16_pool_enabled

Stage 3 of the f16 h-state: h pools SIZED at 2 bytes/element (--ssm-h-dtype f16-pool). Implies ssm_h_fp16_enabled — a narrow pool holding FP32 would be an OOB write, not a mode — which ssm_h_dtype_bits guarantees at the one place the value is decoded.

ssm_h_fp16_enabled

--ssm-h-dtype f16 (legacy ATLAS_SSM_H_FP16).

verify_exact_enabled

--exact-verify given (and h-state is FP32): the MTP-verify pass runs the sequential-decode-exact chain. FALSE by default — without the flag the verify pass runs the WY/chunkwise arms and #435’s spec-on/spec-off output divergence remains. See GdnFlags::verify_exact_active.